Operation Reference

This reference is generated from noema_lab.ops.build_registry() and each operation’s describe() contract.

Operation contracts are the implementation-adjacent source of truth for recipe validation, graph rendering, training inspection, and UI parameter panels.

differentiability.trainable_params is legacy compatibility metadata. The independent training_capabilities values below govern built-in fine-tuning and portable replacement.

channel

channel.bit_boundary

Name: Canonical bit boundary checkpoint

Status: implemented

Inputs:

Name	Kind
bits	channel.payload_bits.numpy, channel.coded_bits.numpy, channel.demod_bits.numpy, channel.bits.numpy

Optional inputs:

None.

Outputs:

Name	Kind
bits	channel.bits.numpy

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
expected_bit_count	integer	no	default 0
label	string	no	default boundary
role	string	no	default channel_boundary

Differentiability (legacy trainable_params): framework=numpy, gradient=stop, trainable_params=False, exportable=False

This fixed-point check validates discrete uint8 bits; it is exact for benchmarking but stops gradients.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy, cpp
dataset_capture	numpy, cpp
differentiable_export	None

Equivalence:

{
  "reason": "Canonical bit boundaries must preserve unpacked uint8 bit values exactly.",
  "type": "exact"
}

Formats:

{
  "artifact": "npz",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "cpp",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "cpp",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

channel.bit_count_match

Name: Channel bit-count consistency check

Status: implemented

Inputs:

Name	Kind
candidate	channel.payload_bits.numpy, channel.coded_bits.numpy, channel.demod_bits.numpy, channel.bits.numpy
reference	channel.payload_bits.numpy, channel.coded_bits.numpy, channel.demod_bits.numpy, channel.bits.numpy

Optional inputs:

None.

Outputs:

Name	Kind
report	metrics.report

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
label	string	no	default channel_io

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

channel.bits_to_indices

Name: Unpack bits into semantic indices

Status: implemented

Inputs:

Name	Kind
bits	channel.payload_bits.numpy, channel.bits.numpy

Optional inputs:

None.

Outputs:

Name	Kind
indices	semantic.indices.numpy

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
invalid_policy	string	no	default mod values mod, clamp

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

channel.bits_to_latents

Name: Unpack payload bits into continuous semantic latents

Status: implemented

Inputs:

Name	Kind
bits	channel.payload_bits.numpy, channel.bits.numpy

Optional inputs:

None.

Outputs:

Name	Kind
latents	semantic.latents.numpy

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
sanitize	boolean	no	default True

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

channel.capacity_oracle_digital_link

Name: Capacity-oracle protected digital AWGN link

Status: implemented

Inputs:

Name	Kind
bits	channel.payload_bits.numpy, channel.coded_bits.numpy, channel.demod_bits.numpy, channel.bits.numpy

Optional inputs:

None.

Outputs:

Name	Kind
bits	channel.bits.numpy

Parameters:

Name	Type	Required	Default / values	Description
channel	string	no	default awgn values awgn
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
label	string	no	default capacity_oracle_awgn
ldpc_rate	number	no	default 0.5
modulation	string	no	default qpsk values qpsk
on_decode_failure	string	no	default gray_image values gray_image, erasure, report_outage
snr_db	number	no	default 12.0

Differentiability (legacy trainable_params): framework=none, gradient=none, trainable_params=False, exportable=False

This is a theoretical protected-digital reference, not a differentiable physical link or implemented LDPC decoder.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

channel.communication_resource_accounting

Name: Communication resource accounting boundary

Status: implemented

Inputs:

Name	Kind
coded_bits	channel.coded_bits.numpy, channel.bits.numpy
framed_bits	channel.payload_bits.numpy, channel.bits.numpy
payload_bits	channel.payload_bits.numpy, channel.bits.numpy
symbols	channel.symbols.complex_numpy

Optional inputs:

None.

Outputs:

Name	Kind
report	metrics.report

Parameters:

Name	Type	Required	Default / values	Description
physical_header_symbol_count	integer	no	default 0
pilot_symbol_count	integer	no	default 0
total_channel_use_count	integer	no	default 0

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

channel.communication_resource_accounting.v2

Name: Communication resource accounting boundary (strict kinds v2)

Status: implemented

Inputs:

Name	Kind
coded_bits	channel.coded_bits.numpy
framed_bits	channel.framed_bits.numpy
payload_bits	channel.payload_bits.numpy
symbols	channel.symbols.complex_numpy

Optional inputs:

None.

Outputs:

Name	Kind
report	metrics.report

Parameters:

Name	Type	Required	Default / values	Description
physical_header_symbol_count	integer	no	default 0
pilot_symbol_count	integer	no	default 0
total_channel_use_count	integer	no	default 0

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

channel.crc32_check

Name: Check CRC32 packets and recover payload bits

Status: implemented

Inputs:

Name	Kind
bits	channel.payload_bits.numpy, channel.bits.numpy, channel.demod_bits.numpy

Optional inputs:

None.

Outputs:

Name	Kind
bits	channel.payload_bits.numpy

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
on_decode_failure	string	no	default gray_image values gray_image, erasure, report_outage

Differentiability (legacy trainable_params): framework=numpy, gradient=stop, trainable_params=False, exportable=False

CRC checking and packet erasure are hard non-differentiable decisions.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

channel.csi_feedback_link

Name: Explicit CSI feedback link

Status: implemented

Inputs:

Name	Kind
feedback_code	channel.csi_feedback_code.numpy

Optional inputs:

None.

Outputs:

Name	Kind
received_code	channel.csi_feedback_received.numpy

Parameters:

Name	Type	Required	Default / values	Description
bits_per_latent	integer	no	default 8
clip_value	number	no	default 1.0
mode	string	no	default ideal_noiseless values ideal_noiseless, uniform_quantized

Differentiability (legacy trainable_params): framework=torch, gradient=surrogate, trainable_params=False, exportable=True

The ideal link is differentiable; fixed-bit quantization requires an externally chosen surrogate such as a straight-through estimator during training.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy
dataset_capture	numpy
differentiable_export	torch

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "npz",
  "tensor": "float32"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "default",
    "runner": "differentiable_export",
    "status": "implemented"
  }
]

channel.identity_decoder

Name: Identity channel decoder

Status: implemented

Inputs:

Name	Kind
coded_bits	channel.coded_bits.numpy, channel.demod_bits.numpy, channel.payload_bits.numpy, channel.bits.numpy

Optional inputs:

None.

Outputs:

Name	Kind
bits	channel.payload_bits.numpy

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

channel.identity_encoder

Name: Identity channel encoder

Status: implemented

Inputs:

Name	Kind
bits	channel.payload_bits.numpy, channel.framed_bits.numpy, channel.bits.numpy

Optional inputs:

None.

Outputs:

Name	Kind
coded_bits	channel.coded_bits.numpy

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

channel.identity_link

Name: Disabled channel identity bit link

Status: implemented

Inputs:

Name	Kind
bits	channel.payload_bits.numpy, channel.coded_bits.numpy, channel.demod_bits.numpy, channel.bits.numpy

Optional inputs:

None.

Outputs:

Name	Kind
bits	channel.bits.numpy

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
label	string	no	default identity_channel

Differentiability (legacy trainable_params): framework=numpy, gradient=stop, trainable_params=False, exportable=False

This is a discrete bitstream bypass. Disabled physical channels for differentiable symbol paths should use channel.identity_symbol_link.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy
dataset_capture	numpy
differentiable_export	None

Equivalence:

{
  "reason": "Bit-perfect bypass materializations must preserve unpacked uint8 bits exactly.",
  "type": "exact"
}

Formats:

{
  "artifact": "npz",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

channel.identity_symbol_link

Name: Disabled channel identity symbol link

Status: implemented

Inputs:

Name	Kind
symbols	channel.symbols.complex_numpy, channel.rx_symbols.complex_numpy

Optional inputs:

None.

Outputs:

Name	Kind
symbols	channel.symbols.complex_numpy

Parameters:

Name	Type	Required	Default / values	Description
label	string	no	default identity_symbol_channel

Differentiability (legacy trainable_params): framework=torch, gradient=full, trainable_params=False, exportable=True

Disabled continuous-symbol channels are identity maps in differentiable export.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy
dataset_capture	numpy
differentiable_export	torch

Equivalence:

{
  "reason": "Identity PHY materializations must return the same continuous symbols.",
  "type": "exact"
}

Formats:

{
  "artifact": "npz",
  "tensor": "torch.Tensor"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "default",
    "runner": "differentiable_export",
    "status": "implemented"
  }
]

channel.indices_to_bits

Name: Pack semantic indices into bits

Status: implemented

Inputs:

Name	Kind
indices	semantic.indices.numpy

Optional inputs:

None.

Outputs:

Name	Kind
bits	channel.payload_bits.numpy

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

channel.jpeg_capacity_oracle

Name: JPEG + ideal-capacity separation

Status: implemented

Inputs:

Name	Kind
images	image.batch.numpy

Optional inputs:

None.

Outputs:

Name	Kind
images	image.batch.numpy

Parameters:

Name	Type	Required	Default / values	Description
channel_model	string	no	default awgn values awgn, slow_rayleigh	awgn uses the average-SNR capacity directly. slow_rayleigh fixes the source rate from average SNR and tests it against one unknown block-fading gain per image.
channel_uses_per_pixel	number	no	default 0.5	Fixed complex channel uses available per spatial source pixel.
maximum_quality	integer	no	default 95
minimum_quality	integer	no	default 1
on_outage	string	no	default gray_image values gray_image, zeros, image_channel_mean
optimize	boolean	no	default False
progressive	boolean	no	default False
seed	integer	no		Optional explicit channel seed. Use the same value on a paired learned recipe to reproduce per-image fading gains.
snr_db	number	no	default 12.0
subsampling	string	no	default 420 values keep, 444, 422, 420

Differentiability (legacy trainable_params): framework=none, gradient=none, trainable_params=False, exportable=False

This is a theoretical separation reference combining classical JPEG with an ideal capacity-achieving channel code.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

Entropy/payload info:

{
  "algorithm": "JPEG Huffman coding",
  "coder": "Huffman",
  "implementation": "Pillow JPEG backend",
  "language": "C/Python",
  "note": "Baseline JPEG comparisons normally use non-progressive JPEG without optimized Huffman-table search."
}

channel.latents_to_bits

Name: Pack continuous semantic latents into payload bits

Status: implemented

Inputs:

Name	Kind
latents	semantic.latents.numpy

Optional inputs:

None.

Outputs:

Name	Kind
bits	channel.payload_bits.numpy

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
dtype	string	no	default float32 values float16, float32

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

channel.nr_ldpc_decoder

Name: 3GPP NR transport-block LDPC soft decoder

Status: implemented

Inputs:

Name	Kind
llr	channel.llr.numpy

Optional inputs:

None.

Outputs:

Name	Kind
bits	channel.payload_bits.numpy
report	metrics.report

Parameters:

Name	Type	Required	Default / values	Description
on_tb_crc_failure	string	no	default zero_fill values zero_fill, keep_estimate, raise

Differentiability (legacy trainable_params): framework=torch, gradient=stop, trainable_params=False, exportable=False

The publication receiver returns hard bits and CRC decisions.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	sionna
dataset_capture	sionna
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "sionna",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "sionna",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

channel.nr_ldpc_encoder

Name: 3GPP NR transport-block LDPC encoder

Status: implemented

Inputs:

Name	Kind
bits	channel.payload_bits.numpy, channel.framed_bits.numpy, channel.bits.numpy

Optional inputs:

None.

Outputs:

Name	Kind
coded_bits	channel.coded_bits.numpy

Parameters:

Name	Type	Required	Default / values	Description
channel_type	string	no	default PUSCH values PUSCH, PDSCH
codeword_index	integer	no	default 0 values 0, 1
n_id	integer	no	default 1
n_rnti	integer	no	default 1
num_bits_per_symbol	integer	no	default 2 values 1, 2, 4, 6, 8
num_bp_iter	integer	no	default 20
num_layers	integer	no	default 1
target_coderate	number	no	default 0.5
transport_block_size_bits	integer	no	default 16000

Differentiability (legacy trainable_params): framework=torch, gradient=stop, trainable_params=False, exportable=False

CRC attachment, segmentation, rate matching, and hard bit coding are discrete.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	sionna
dataset_capture	sionna
differentiable_export	None

Equivalence:

{
  "reason": "Publication use requires independent TS 38.212 vectors or a second standards implementation; internal round trips alone are not conformance.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "sionna",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "sionna",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

channel.packetize_crc32

Name: Packetize payload bits with CRC32

Status: implemented

Inputs:

Name	Kind
bits	channel.payload_bits.numpy, channel.bits.numpy

Optional inputs:

None.

Outputs:

Name	Kind
bits	channel.payload_bits.numpy

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
packet_payload_bits	integer	no	default 4096	Payload bits protected by one CRC32. The final packet is padded but its valid length is checked.

Differentiability (legacy trainable_params): framework=numpy, gradient=stop, trainable_params=False, exportable=False

CRC packetization is a hard digital transport operation.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

channel.packetize_crc32.v2

Name: Packetize payload bits with CRC32 (framed-bit contract v2)

Status: implemented

Inputs:

Name	Kind
bits	channel.payload_bits.numpy

Optional inputs:

None.

Outputs:

Name	Kind
bits	channel.framed_bits.numpy

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
packet_payload_bits	integer	no	default 4096	Payload bits protected by one CRC32. The final packet is padded but its valid length is checked.

Differentiability (legacy trainable_params): framework=numpy, gradient=stop, trainable_params=False, exportable=False

CRC packetization is a hard digital transport operation.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

channel.payload_passthrough_decoder

Name: Payload bitstream pass-through decoder

Status: implemented

Inputs:

Name	Kind
bits	channel.payload_bits.numpy, channel.bits.numpy

Optional inputs:

None.

Outputs:

Name	Kind
bits	channel.payload_bits.numpy

Parameters:

None.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

channel.payload_passthrough_encoder

Name: Payload bitstream pass-through encoder

Status: implemented

Inputs:

Name	Kind
bits	channel.payload_bits.numpy, channel.bits.numpy

Optional inputs:

None.

Outputs:

Name	Kind
bits	channel.payload_bits.numpy

Parameters:

None.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

channel.repetition_decoder

Name: Repetition channel decoder

Status: implemented

Inputs:

Name	Kind
coded_bits	channel.coded_bits.numpy, channel.demod_bits.numpy, channel.bits.numpy

Optional inputs:

None.

Outputs:

Name	Kind
bits	channel.payload_bits.numpy

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
factor	integer	no	default 0

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

channel.repetition_encoder

Name: Repetition channel encoder

Status: implemented

Inputs:

Name	Kind
bits	channel.payload_bits.numpy, channel.framed_bits.numpy, channel.bits.numpy

Optional inputs:

None.

Outputs:

Name	Kind
coded_bits	channel.coded_bits.numpy

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
factor	integer	no	default 3

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

channel.symbol_boundary

Name: Canonical complex-symbol boundary checkpoint

Status: implemented

Inputs:

Name	Kind
symbols	channel.symbols.complex_numpy, channel.rx_symbols.complex_numpy

Optional inputs:

None.

Outputs:

Name	Kind
symbols	channel.symbols.complex_numpy

Parameters:

Name	Type	Required	Default / values	Description
expected_symbol_count	integer	no	default 0
label	string	no	default symbol_boundary
role	string	no	default channel_symbol_boundary

Differentiability (legacy trainable_params): framework=torch, gradient=full, trainable_params=False, exportable=True

Differentiable export treats this fixed-point check as a typed identity for continuous symbols; benchmark execution validates and serializes complex64 symbols.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy
dataset_capture	numpy
differentiable_export	torch

Equivalence:

{
  "reason": "Symbol boundaries are identity checks; numeric implementations must preserve complex symbol values.",
  "tolerance": {
    "atol": 0.0,
    "rtol": 0.0
  },
  "type": "numerical"
}

Formats:

{
  "artifact": "npz",
  "tensor": "torch.Tensor"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "default",
    "runner": "differentiable_export",
    "status": "implemented"
  }
]

channel.symbol_count_match

Name: Channel symbol-count consistency check

Status: implemented

Inputs:

Name	Kind
candidate	channel.symbols.complex_numpy, channel.rx_symbols.complex_numpy
reference	channel.symbols.complex_numpy, channel.rx_symbols.complex_numpy

Optional inputs:

None.

Outputs:

Name	Kind
report	metrics.report

Parameters:

Name	Type	Required	Default / values	Description
label	string	no	default symbol_channel_io

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

channel.symbol_power_identity

Name: Transmit power disabled pass-through

Status: implemented

Inputs:

Name	Kind
symbols	channel.symbols.complex_numpy, channel.rx_symbols.complex_numpy

Optional inputs:

None.

Outputs:

Name	Kind
symbols	channel.symbols.complex_numpy

Parameters:

Name	Type	Required	Default / values	Description
label	string	no	default tx_power_off
power_unit	string	no	default normalized values normalized, mW

Differentiability (legacy trainable_params): framework=torch, gradient=full, trainable_params=False, exportable=True

Disabled TX power is a continuous-symbol identity block; it preserves the PHY skeleton and does not change transmit power.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy
dataset_capture	numpy
differentiable_export	torch

Equivalence:

{
  "reason": "The disabled TX power block must pass canonical complex symbols through unchanged.",
  "type": "exact"
}

Formats:

{
  "artifact": "npz",
  "tensor": "torch.Tensor"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "numpy_symbol_identity",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "numpy_symbol_identity",
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "torch_identity",
    "runner": "differentiable_export",
    "status": "implemented"
  }
]

channel.symbol_power_normalize

Name: Transmit symbol power normalization

Status: implemented

Inputs:

Name	Kind
symbols	channel.symbols.complex_numpy, channel.rx_symbols.complex_numpy

Optional inputs:

None.

Outputs:

Name	Kind
symbols	channel.symbols.complex_numpy

Parameters:

Name	Type	Required	Default / values	Description
eps	number	no	default 1e-12
label	string	no	default tx_power_normalize
normalization_scope	string	no	default source_item values source_item, global	Normalize each declared source item independently, or normalize the complete tensor/stream as one global batch.
target_power	number	no	default 1.0	Normalized average complex-symbol power target applied by this normalizer; this is not a value in watts unless a physical link-budget conversion is defined.

Differentiability (legacy trainable_params): framework=torch, gradient=full, trainable_params=False, exportable=True

Continuous-symbol normalization is a differentiable scale operation that fixes average transmit power before the physical channel.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy
dataset_capture	numpy
differentiable_export	torch

Equivalence:

{
  "reason": "NumPy and Torch materializations should produce the same average-power-normalized symbols within floating-point tolerance.",
  "tolerance": {
    "atol": 1e-06,
    "rtol": 1e-05
  },
  "type": "numerical"
}

Formats:

{
  "artifact": "npz",
  "tensor": "torch.Tensor"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "numpy_source_item_power_normalizer",
    "parameter_bindings": {
      "normalization_scope": "source_item"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "numpy_global_power_normalizer",
    "parameter_bindings": {
      "normalization_scope": "global"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "numpy_source_item_power_normalizer",
    "parameter_bindings": {
      "normalization_scope": "source_item"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "numpy_global_power_normalizer",
    "parameter_bindings": {
      "normalization_scope": "global"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "torch_source_item_power_normalizer",
    "parameter_bindings": {
      "normalization_scope": "source_item"
    },
    "runner": "differentiable_export",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "torch_global_power_normalizer",
    "parameter_bindings": {
      "normalization_scope": "global"
    },
    "runner": "differentiable_export",
    "status": "implemented"
  }
]

demodulation

demodulation.digital_demodulate

Name: Digital symbol-to-bit demodulator

Status: implemented

Inputs:

Name	Kind
rx_symbols	channel.symbols.complex_numpy, channel.rx_symbols.complex_numpy

Optional inputs:

Name	Kind
allocation	channel.power_allocation.numpy
channel_state	channel.ofdm_channel_state.numpy

Outputs:

Name	Kind
bits	channel.demod_bits.numpy
llr	channel.llr.numpy

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
modulation	string	no	default auto values auto, bpsk, qpsk, qam16

Differentiability (legacy trainable_params): framework=numpy, gradient=stop, trainable_params=False, exportable=False

Hard demodulation decisions stop gradients; use soft/differentiable demodulation or dataset capture mode for neural receiver training.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy, cpp
dataset_capture	numpy, cpp
differentiable_export	None

Equivalence:

{
  "reason": "Demapper materializations should produce matching hard bits and comparable LLR-like reliability values for the same symbols.",
  "tolerance": {
    "atol": 1e-07,
    "rtol": 1e-06
  },
  "type": "numerical"
}

Formats:

{
  "artifact": "npz",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "numpy_auto_demodulator",
    "parameter_bindings": {
      "modulation": "auto"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "numpy_bpsk_demodulator",
    "parameter_bindings": {
      "modulation": "bpsk"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "numpy_qpsk_demodulator",
    "parameter_bindings": {
      "modulation": "qpsk"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "numpy_qam16_demodulator",
    "parameter_bindings": {
      "modulation": "qam16"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "cpp",
    "implementation": "cpp_bpsk_demodulator",
    "parameter_bindings": {
      "modulation": "bpsk"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "cpp",
    "implementation": "cpp_qpsk_demodulator",
    "parameter_bindings": {
      "modulation": "qpsk"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "numpy_auto_demodulator",
    "parameter_bindings": {
      "modulation": "auto"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "numpy_bpsk_demodulator",
    "parameter_bindings": {
      "modulation": "bpsk"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "numpy_qpsk_demodulator",
    "parameter_bindings": {
      "modulation": "qpsk"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "numpy_qam16_demodulator",
    "parameter_bindings": {
      "modulation": "qam16"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "cpp",
    "implementation": "cpp_bpsk_demodulator",
    "parameter_bindings": {
      "modulation": "bpsk"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "cpp",
    "implementation": "cpp_qpsk_demodulator",
    "parameter_bindings": {
      "modulation": "qpsk"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

demodulation.identity_demodulate

Name: Identity PHY symbol-to-bit mapper

Status: implemented

Inputs:

Name	Kind
rx_symbols	channel.symbols.complex_numpy, channel.rx_symbols.complex_numpy

Optional inputs:

None.

Outputs:

Name	Kind
bits	channel.demod_bits.numpy
llr	channel.llr.numpy

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
label	string	no	default identity_phy_demodulator

Differentiability (legacy trainable_params): framework=torch, gradient=surrogate, trainable_params=False, exportable=True

Artifact execution thresholds 0/1 identity PHY symbols back to bits. Differentiable export may keep a continuous soft/logit materialization until a loss, but hard bit decisions are surrogate-only.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy
dataset_capture	numpy
differentiable_export	torch

Equivalence:

{
  "reason": "When paired with identity modulation and identity channel, demodulation recovers the original uint8 bit vector exactly.",
  "type": "exact"
}

Formats:

{
  "artifact": "npz",
  "tensor": "torch.Tensor"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "default",
    "runner": "differentiable_export",
    "status": "implemented"
  }
]

demodulation.neural_receiver_adapter

Name: Neural receiver adapter

Status: implemented

Inputs:

Name	Kind
rx_symbols	channel.rx_symbols.complex_numpy

Optional inputs:

None.

Outputs:

Name	Kind
bits	channel.demod_bits.numpy
llr	channel.llr.numpy

Parameters:

Name	Type	Required	Default / values	Description
artifact_entrypoint	string	no	default neural_receiver	Entrypoint in the registered trained artifact used for inference.
artifact_manifest_path	string	no	default ``	Registered schema-v2 trained artifact implementing the neural-receiver slot ABI.
artifact_package_sha256	string	no	default ``	Registry-independent digest of the complete trained-artifact package bound into the execution plan.
checkpoint_path	string	no	default ``
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
mode	string	no	default reference_qpsk values reference_qpsk, oracle_frontend_calibrated, linear_npz, learned_artifact
modulation	string	no	default auto values auto, qpsk

Differentiability (legacy trainable_params): framework=torch, gradient=surrogate, trainable_params=True, exportable=True

Checkpoint-backed neural receivers can be trained/exported; artifact benchmark mode records hard decisions and treats them as gradient-stopping evidence.

Training capabilities: built_in_fine_tuning=False, portable_replacement=True

Portable trained-artifact ABI:

{
  "binding_params": {
    "artifact_entrypoint": "neural_receiver",
    "artifact_manifest_path": "trained_artifact.yaml",
    "mode": "learned_artifact",
    "modulation": "qpsk"
  },
  "component_id": "receiver",
  "component_role": "neural_receiver",
  "entrypoint_id": "neural_receiver",
  "inputs": {
    "rx_symbols_ri": {
      "dtype": "float32",
      "shape": [
        "symbol",
        2
      ]
    }
  },
  "outputs": {
    "bit_llr": {
      "dtype": "float32",
      "shape": [
        "symbol",
        2
      ]
    }
  },
  "required_operation_inputs": [
    "rx_symbols"
  ]
}

Backends:

Runner	Backends
benchmark_run	numpy, onnxruntime
dataset_capture	numpy, onnxruntime
differentiable_export	torch

Equivalence:

{
  "reason": "External neural receiver checkpoints are compared by declared BER/BLER behavior rather than exact implementation internals.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "npz",
  "checkpoint": "npz|onnx",
  "tensor": "torch.Tensor"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "reference_qpsk_receiver",
    "notes": "Built-in reference QPSK receiver for smoke benchmarks.",
    "parameter_bindings": {
      "mode": "reference_qpsk"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "calibrated_iq_oracle_receiver",
    "notes": "Diagnostic oracle that inverts the simulated fixed I/Q front-end transform before QPSK demapping.",
    "parameter_bindings": {
      "mode": "oracle_frontend_calibrated"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "linear_npz_receiver",
    "notes": "Tiny linear NPZ checkpoint adapter for smoke benchmarks.",
    "parameter_bindings": {
      "mode": "linear_npz"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "reference_qpsk_receiver",
    "notes": "Produces demodulated bits and LLR-like logits while preserving rx-symbol capture compatibility.",
    "parameter_bindings": {
      "mode": "reference_qpsk"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "calibrated_iq_oracle_receiver",
    "notes": "Uses simulation-only front-end calibration metadata for a diagnostic upper-bound receiver.",
    "parameter_bindings": {
      "mode": "oracle_frontend_calibrated"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "linear_npz_receiver",
    "notes": "Runs the linear NPZ checkpoint adapter while preserving capture compatibility.",
    "parameter_bindings": {
      "mode": "linear_npz"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "onnxruntime",
    "implementation": "portable_trained_artifact_runtime",
    "notes": "Runs a schema-v2, hash-pinned neural-receiver artifact through its operation-owned tensor ABI.",
    "parameter_bindings": {
      "mode": "learned_artifact"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "onnxruntime",
    "implementation": "portable_trained_artifact_runtime",
    "notes": "The same returned artifact can be used in ordinary capture recipes without a task-specific runner.",
    "parameter_bindings": {
      "mode": "learned_artifact"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "neural_receiver_training_endpoint",
    "notes": "Training-contract export exposes this typed receiver slot; the researcher supplies the trainable module.",
    "runner": "differentiable_export",
    "status": "implemented"
  }
]

demodulation.phase_tracking_receiver_adapter

Name: Pilot-aided QPSK phase-tracking receiver

Status: implemented

Inputs:

Name	Kind
pilot_context	channel.qpsk_pilot_context.numpy
rx_symbols	channel.rx_symbols.complex_numpy

Optional inputs:

Name	Kind
phase_truth	channel.carrier_phase_truth.numpy

Outputs:

Name	Kind
bits	channel.demod_bits.numpy
diagnostics	receiver.phase_tracking_diagnostics.numpy
llr	channel.llr.numpy

Parameters:

Name	Type	Required	Default / values	Description
artifact_entrypoint	string	no	default phase_tracking_receiver
artifact_manifest_path	string	no	default ``	Registered schema-v2 artifact implementing the packet-context phase-tracking receiver ABI.
artifact_package_sha256	string	no	default ``
mode	string	no	default pilot_smoothing values uncompensated, pilot_interpolation, pilot_smoothing, decision_directed_pll, oracle, learned_artifact
pilot_smoothing_nearest_pilots	integer	no	default 5	Number of nearest public pilots used by the noncausal local-linear packet smoother.
pll_alpha	number	no	default 0.12
pll_beta	number	no	default 0.005

Differentiability (legacy trainable_params): framework=torch, gradient=surrogate, trainable_params=True, exportable=True

The adapter exposes a packet-context portable receiver slot while classical modes remain reproducible baselines.

Training capabilities: built_in_fine_tuning=False, portable_replacement=True

Portable trained-artifact ABI:

{
  "binding_params": {
    "artifact_entrypoint": "phase_tracking_receiver",
    "artifact_manifest_path": "trained_artifact.yaml",
    "mode": "learned_artifact"
  },
  "component_id": "receiver",
  "component_role": "phase_tracking_receiver",
  "entrypoint_id": "phase_tracking_receiver",
  "inputs": {
    "receiver_features_v3": {
      "dtype": "float32",
      "shape": [
        "packet",
        "frame_symbol",
        11
      ]
    }
  },
  "outputs": {
    "residual_phase_rad": {
      "dtype": "float32",
      "shape": [
        "packet",
        "frame_symbol"
      ]
    }
  },
  "required_operation_inputs": [
    "rx_symbols",
    "pilot_context"
  ]
}

Backends:

Runner	Backends
benchmark_run	numpy, onnxruntime
dataset_capture	numpy, onnxruntime
differentiable_export	torch

Equivalence:

{
  "reason": "Phase-tracking receivers are compared by paired BER/BLER behavior, not sample-identical internal estimates.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "npz",
  "checkpoint": "onnx",
  "tensor": "torch.Tensor"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "uncompensated_phase_tracking_receiver",
    "parameter_bindings": {
      "mode": "uncompensated"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "pilot_interpolation_phase_tracking_receiver",
    "parameter_bindings": {
      "mode": "pilot_interpolation"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "pilot_smoothing_phase_tracking_receiver",
    "parameter_bindings": {
      "mode": "pilot_smoothing"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "decision_directed_pll_phase_tracking_receiver",
    "parameter_bindings": {
      "mode": "decision_directed_pll"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "oracle_phase_tracking_receiver",
    "parameter_bindings": {
      "mode": "oracle"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "uncompensated_phase_tracking_receiver",
    "parameter_bindings": {
      "mode": "uncompensated"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "pilot_interpolation_phase_tracking_receiver",
    "parameter_bindings": {
      "mode": "pilot_interpolation"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "pilot_smoothing_phase_tracking_receiver",
    "parameter_bindings": {
      "mode": "pilot_smoothing"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "decision_directed_pll_phase_tracking_receiver",
    "parameter_bindings": {
      "mode": "decision_directed_pll"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "oracle_phase_tracking_receiver",
    "parameter_bindings": {
      "mode": "oracle"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "onnxruntime",
    "implementation": "portable_packet_context_receiver",
    "notes": "Simulator phase truth is not passed to the learned entrypoint.",
    "parameter_bindings": {
      "mode": "learned_artifact"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "onnxruntime",
    "implementation": "portable_packet_context_receiver",
    "notes": "Simulator phase truth is not passed to the learned entrypoint.",
    "parameter_bindings": {
      "mode": "learned_artifact"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "phase_tracking_receiver_training_endpoint",
    "runner": "differentiable_export",
    "status": "implemented"
  }
]

foundation

foundation.clip_image_embed

Name: CLIP-style image embedding adapter

Status: implemented

Inputs:

Name	Kind
images	image.batch.numpy

Optional inputs:

None.

Outputs:

Name	Kind
embeddings	foundation.embedding.numpy

Parameters:

Name	Type	Required	Default / values	Description
backend	string	no	default local_semantic values local_semantic, local_color_histogram, transformers_clip
bins	integer	no	default 16
device	string	no	default cpu
dimensions	integer	no	default 8
model_id	string	no	default openai/clip-vit-base-patch32
model_revision	string	no	default ``	Full immutable Hugging Face commit. The built-in default model resolves to its pinned commit; other remote models require this explicitly.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

foundation.clip_retrieval_rank

Name: CLIP retrieval ranker

Status: implemented

Inputs:

Name	Kind
image_embeddings	foundation.embedding.numpy, vision.embedding.clip.numpy, multimodal.embedding.numpy
targets	retrieval.targets.json
text_embeddings	foundation.embedding.numpy, vision.embedding.clip.numpy, multimodal.embedding.numpy

Optional inputs:

None.

Outputs:

Name	Kind
rankings	retrieval.rankings.json

Parameters:

Name	Type	Required	Default / values	Description
normalize	boolean	no	default True
top_k	integer	no	default 10

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

foundation.clip_text_embed

Name: CLIP-style text embedding adapter

Status: implemented

Inputs:

Name	Kind
texts	text.batch.json

Optional inputs:

None.

Outputs:

Name	Kind
embeddings	foundation.embedding.numpy

Parameters:

Name	Type	Required	Default / values	Description
backend	string	no	default local_semantic values local_semantic, local_hash, transformers_clip
device	string	no	default cpu
dimensions	integer	no	default 128
model_id	string	no	default openai/clip-vit-base-patch32
model_revision	string	no	default ``	Full immutable Hugging Face commit. The built-in default model resolves to its pinned commit; other remote models require this explicitly.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

foundation.diffusion_state_to_image

Name: Diffusion SemanticState-to-image adapter

Status: implemented

Inputs:

Name	Kind
state	semantic.state.json

Optional inputs:

None.

Outputs:

Name	Kind
images	image.batch.numpy

Parameters:

Name	Type	Required	Default / values	Description
backend	string	no	default diffusers values diffusers
device	string	no	default cpu
disable_safety_checker	boolean	no	default False
guidance_scale	number	no	default 7.5
height	integer	no	default 512
model_id	string	no	default segmind/tiny-sd
model_revision	string	no	default ``	Full immutable Hugging Face commit. The built-in default model resolves to its pinned commit; other remote models require this explicitly.
negative_prompt	string	no	default low quality, blurry, abstract, distorted, text, watermark
num_inference_steps	integer	no	default 20
prompt_template	string	no	default A realistic photograph: {prompt}
use_safetensors	boolean | null	no	default None
width	integer	no	default 512

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

foundation.knowledge_base

Name: Foundation knowledge-base source

Status: implemented

Inputs:

None.

Optional inputs:

None.

Outputs:

Name	Kind
kb	foundation.kb.json

Parameters:

Name	Type	Required	Default / values	Description
facts_json	string	no	default ``	Optional JSON list of facts. Used when kb_id is inline_json; appended for semantic_text_smoke_kb.
kb_id	string	no	default semantic_text_smoke_kb values semantic_text_smoke_kb, empty, inline_json
max_builtin_concepts	integer	no	default 12

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

foundation.sam_segment

Name: SAM-style image segmentation adapter

Status: implemented

Inputs:

Name	Kind
images	image.batch.numpy

Optional inputs:

None.

Outputs:

Name	Kind
state	semantic.state.json

Parameters:

Name	Type	Required	Default / values	Description
backend	string	no	default local_grid values local_grid
grid_size	integer	no	default 2

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

foundation.semantic_state_ground

Name: Ground SemanticState against a knowledge base

Status: implemented

Inputs:

Name	Kind
kb	foundation.kb.json
state	semantic.state.json

Optional inputs:

None.

Outputs:

Name	Kind
state	semantic.state.json

Parameters:

Name	Type	Required	Default / values	Description
max_matches_per_state	integer	no	default 12
min_token_overlap	integer	no	default 1

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

foundation.semantic_state_payload_decode

Name: SemanticState payload decoder

Status: implemented

Inputs:

Name	Kind
bits	channel.payload_bits.numpy, channel.bits.numpy

Optional inputs:

None.

Outputs:

Name	Kind
state	semantic.state.json

Parameters:

Name	Type	Required	Default / values	Description
on_error	string	no	default replace values replace, fail

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

foundation.semantic_state_payload_encode

Name: SemanticState payload encoder

Status: implemented

Inputs:

Name	Kind
state	semantic.state.json

Optional inputs:

None.

Outputs:

Name	Kind
bits	channel.payload_bits.numpy

Parameters:

Name	Type	Required	Default / values	Description
payload_format	string	no	default semantic_state_json_utf8 values semantic_state_json_utf8

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

foundation.semantic_state_to_text

Name: SemanticState to text generator

Status: implemented

Inputs:

Name	Kind
kb	foundation.kb.json
state	semantic.state.json

Optional inputs:

None.

Outputs:

Name	Kind
texts	text.batch.json

Parameters:

Name	Type	Required	Default / values	Description
cache_dir	string	no	default ``
device	string	no	default cpu
generator	string	no	default local_template values local_template, kb_reconstruct, masked_lm
model_id	string	no	default distilbert/distilbert-base-uncased
model_revision	string	no	default ``	Full immutable Hugging Face commit. The built-in default model resolves to its pinned commit; other remote models require this explicitly.
prefer_source_text	boolean	no	default True
top_k	integer	no	default 1

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

foundation.text_mask_repair

Name: Text receiver mask repair

Status: implemented

Inputs:

Name	Kind
texts	text.batch.json

Optional inputs:

None.

Outputs:

Name	Kind
texts	text.batch.json

Parameters:

Name	Type	Required	Default / values	Description
cache_dir	string	no	default ``
device	string	no	default cpu
generator	string	no	default masked_lm values local_template, masked_lm
model_id	string	no	default distilbert/distilbert-base-uncased
model_revision	string	no	default ``	Full immutable Hugging Face commit. The built-in default model resolves to its pinned commit; other remote models require this explicitly.
top_k	integer	no	default 1

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

foundation.text_semantic_state_encode

Name: Text to SemanticState encoder

Status: implemented

Inputs:

Name	Kind
texts	text.batch.json

Optional inputs:

None.

Outputs:

Name	Kind
state	semantic.state.json

Parameters:

Name	Type	Required	Default / values	Description
extractor	string	no	default local_rules values local_rules
include_text	boolean	no	default False
max_concepts	integer	no	default 16

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

foundation.vlm_image_to_state

Name: VLM-style image SemanticState adapter

Status: implemented

Inputs:

Name	Kind
images	image.batch.numpy

Optional inputs:

None.

Outputs:

Name	Kind
state	semantic.state.json

Parameters:

Name	Type	Required	Default / values	Description
backend	string	no	default local_image_stats values local_image_stats
model_id	string	no	default ``

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

foundation.vqa_answer_from_packet

Name: VQA answer from semantic packet

Status: implemented

Inputs:

Name	Kind
packet	vqa.semantic_packet.json

Optional inputs:

None.

Outputs:

Name	Kind
answers	vqa.answers.json

Parameters:

Name	Type	Required	Default / values	Description
receiver	string	no	default selected_label values selected_label

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

foundation.vqa_payload_decode

Name: VQA semantic packet payload decoder

Status: implemented

Inputs:

Name	Kind
bits	channel.payload_bits.numpy, channel.bits.numpy

Optional inputs:

None.

Outputs:

Name	Kind
packet	vqa.semantic_packet.json

Parameters:

Name	Type	Required	Default / values	Description
on_error	string	no	default fail values fail, replace

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

foundation.vqa_payload_encode

Name: VQA semantic packet payload encoder

Status: implemented

Inputs:

Name	Kind
packet	vqa.semantic_packet.json

Optional inputs:

None.

Outputs:

Name	Kind
bits	channel.payload_bits.numpy

Parameters:

Name	Type	Required	Default / values	Description
payload_format	string	no	default json_utf8 values json_utf8

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

foundation.vqa_semantic_select

Name: VQA semantic region selector

Status: implemented

Inputs:

Name	Kind
images	image.batch.numpy
questions	vqa.questions.json

Optional inputs:

None.

Outputs:

Name	Kind
detections	vision.detections.json
packet	vqa.semantic_packet.json

Parameters:

Name	Type	Required	Default / values	Description
selector	string	no	default local_rules values local_rules

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

foundation.vqa_transformers_answer

Name: Pretrained Transformers VQA answerer

Status: implemented

Inputs:

Name	Kind
images	image.batch.numpy
questions	vqa.questions.json

Optional inputs:

None.

Outputs:

Name	Kind
answers	vqa.answers.json

Parameters:

Name	Type	Required	Default / values	Description
device	string	no	default cpu	cpu or cuda device string. The operation uses transformers.pipeline device mapping.
model_id	string	no	default dandelin/vilt-b32-finetuned-vqa	Hugging Face model id for a visual-question-answering pipeline. Practical defaults include dandelin/vilt-b32-finetuned-vqa and Salesforce/blip-vqa-base.
model_revision	string	no	default ``	Full immutable Hugging Face commit. The built-in default model resolves to its pinned commit; other remote models require this explicitly.
top_k	integer	no	default 1

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

foundation.yolo_detect

Name: YOLO object detector

Status: implemented

Inputs:

Name	Kind
images	image.batch.numpy

Optional inputs:

None.

Outputs:

Name	Kind
detections	vision.detections.json

Parameters:

Name	Type	Required	Default / values	Description
confidence	number	no	default 0.25
device	string	no	default cpu
expected_model_sha256	string	no	default ``	Required for a remote model URL; known built-in asset names use release-pinned hashes.
iou	number	no	default 0.7
model_id	string	no	default yolo11n.pt

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

foundation.yolo_segment

Name: YOLO instance segmenter

Status: implemented

Inputs:

Name	Kind
images	image.batch.numpy

Optional inputs:

None.

Outputs:

Name	Kind
segmentation	vision.segmentation_mask.numpy

Parameters:

Name	Type	Required	Default / values	Description
confidence	number	no	default 0.25
device	string	no	default cpu
expected_model_sha256	string	no	default ``	Required for a remote model URL; known built-in asset names use release-pinned hashes.
iou	number	no	default 0.7
mask_threshold	number	no	default 0.5
model_id	string	no	default yolo11n-seg.pt

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

hardware

hardware.receiver_iq_imbalance

Name: Receiver I/Q front-end impairment

Status: implemented

Inputs:

Name	Kind
rx_symbols	channel.rx_symbols.complex_numpy

Optional inputs:

None.

Outputs:

Name	Kind
rx_symbols	channel.rx_symbols.complex_numpy

Parameters:

Name	Type	Required	Default / values	Description
dc_offset_i	number	no	default 0.18
dc_offset_q	number	no	default -0.12
gain_imbalance_db	number	no	default 5.0	I-to-Q amplitude-gain ratio in dB.
phase_offset_deg	number	no	default 20.0
quadrature_error_deg	number	no	default 12.0

Differentiability (legacy trainable_params): framework=torch, gradient=full, trainable_params=False, exportable=True

The fixed affine I/Q transform is differentiable, while its simulated calibration parameters remain frozen scenario settings.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy
dataset_capture	numpy
differentiable_export	torch

Equivalence:

{
  "reason": "All materializations implement the same frozen real 2x2 I/Q transform and DC offset.",
  "tolerance": {
    "atol": 1e-06,
    "rtol": 1e-06
  },
  "type": "numerical"
}

Formats:

{
  "artifact": "npz",
  "tensor": "torch.Tensor"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "fixed_affine_receiver_iq_impairment",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "fixed_affine_receiver_iq_impairment",
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "fixed_affine_receiver_iq_impairment",
    "runner": "differentiable_export",
    "status": "implemented"
  }
]

metrics

metrics.aoa_estimation

Name: AoA-estimation metrics

Status: implemented

Inputs:

Name	Kind
estimate	ai_phy.aoa_estimate.numpy
problem	ai_phy.aoa_problem.numpy, ai_phy.aoa_truth.numpy

Optional inputs:

None.

Outputs:

Name	Kind
report	metrics.report

Parameters:

None.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

metrics.beamforming

Name: Beamforming metrics

Status: implemented

Inputs:

Name	Kind
decision	ai_phy.beamforming_decision.numpy
problem	ai_phy.beamforming_problem.numpy

Optional inputs:

None.

Outputs:

Name	Kind
report	metrics.report

Parameters:

None.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

metrics.bit_error_rate

Name: Bit error rate

Status: implemented

Inputs:

Name	Kind
candidate	channel.payload_bits.numpy, channel.coded_bits.numpy, channel.demod_bits.numpy, channel.bits.numpy
reference	channel.payload_bits.numpy, channel.coded_bits.numpy, channel.demod_bits.numpy, channel.bits.numpy

Optional inputs:

None.

Outputs:

Name	Kind
report	metrics.report

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
label	string	no	default ber

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

metrics.block_error_rate

Name: Block error rate

Status: implemented

Inputs:

Name	Kind
candidate	channel.payload_bits.numpy, channel.coded_bits.numpy, channel.demod_bits.numpy, channel.bits.numpy
reference	channel.payload_bits.numpy, channel.coded_bits.numpy, channel.demod_bits.numpy, channel.bits.numpy

Optional inputs:

None.

Outputs:

Name	Kind
report	metrics.report

Parameters:

Name	Type	Required	Default / values	Description
block_size	integer	no	default 1024
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
label	string	no	default bler

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

metrics.captioning

Name: Image captioning task metrics

Status: implemented

Inputs:

Name	Kind
candidate	text.batch.json, text.caption.json
reference	text.batch.json, text.caption.json

Optional inputs:

None.

Outputs:

Name	Kind
report	metrics.report

Parameters:

None.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

metrics.channel_estimation

Name: Channel-estimation metrics

Status: implemented

Inputs:

Name	Kind
estimate	ai_phy.channel_estimate.numpy
problem	ai_phy.channel_estimation_problem.numpy, ai_phy.channel_estimation_truth.numpy

Optional inputs:

None.

Outputs:

Name	Kind
report	metrics.report

Parameters:

None.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

metrics.classification

Name: Classification task metrics

Status: implemented

Inputs:

Name	Kind
candidate	task.predictions.json, task.labels.json
reference	task.labels.json

Optional inputs:

None.

Outputs:

Name	Kind
report	metrics.report

Parameters:

None.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

metrics.csi_feedback

Name: CSI feedback and downlink metrics

Status: implemented

Inputs:

Name	Kind
precoder	channel.miso_ofdm_precoder.numpy
reconstruction	channel.miso_ofdm_csi_reconstruction.numpy
true_csi	channel.miso_ofdm_csi.numpy

Optional inputs:

None.

Outputs:

Name	Kind
report	metrics.report

Parameters:

None.

Differentiability (legacy trainable_params): framework=torch, gradient=full, trainable_params=False, exportable=True

NMSE and reconstructed-CSI MRT achievable rate have direct differentiable tensor forms.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy
dataset_capture	numpy
differentiable_export	torch

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "json",
  "tensor": "structured metrics"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "default",
    "runner": "differentiable_export",
    "status": "implemented"
  }
]

metrics.detection

Name: Object detection task metrics

Status: implemented

Inputs:

Name	Kind
candidate	vision.detections.json
reference	vision.detections.json

Optional inputs:

None.

Outputs:

Name	Kind
report	metrics.report

Parameters:

Name	Type	Required	Default / values	Description
confidence_threshold	number	no	default 0.0
iou_threshold	number	no	default 0.5

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

metrics.embedding_similarity

Name: Paired embedding similarity metrics

Status: implemented

Inputs:

Name	Kind
candidate	foundation.embedding.numpy, vision.embedding.clip.numpy, multimodal.embedding.numpy
reference	foundation.embedding.numpy, vision.embedding.clip.numpy, multimodal.embedding.numpy

Optional inputs:

None.

Outputs:

Name	Kind
report	metrics.report

Parameters:

Name	Type	Required	Default / values	Description
label	string	no	default embedding values embedding, generation.text_image_clip, generation.image_image_clip
normalize	boolean	no	default True

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

metrics.external_classification

Name: External classification metric

Status: implemented

Inputs:

Name	Kind
candidate	task.predictions.json, task.labels.json
reference	task.labels.json

Optional inputs:

None.

Outputs:

Name	Kind
report	metrics.report

Parameters:

Name	Type	Required	Default / values	Description
call_style	string	no	default dict values dict, params, none
callable	string	no	default ``
module	string	no	default ``
path	string	no	default ``

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

metrics.image_delivery_status

Name: Image delivery and denominator status

Status: implemented

Inputs:

Name	Kind
images	image.batch.numpy

Optional inputs:

None.

Outputs:

Name	Kind
report	metrics.report

Parameters:

Name	Type	Required	Default / values	Description
on_decode_failure	string	no	default report_outage values report_outage

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

metrics.image_reconstruction

Name: Image reconstruction metrics

Status: implemented

Inputs:

Name	Kind
reconstruction	image.batch.numpy
reference	image.batch.numpy

Optional inputs:

None.

Outputs:

Name	Kind
report	metrics.report

Parameters:

Name	Type	Required	Default / values	Description
data_range	number	no	default 0.0	Maximum minus minimum pixel value used by PSNR; 0 infers the canonical range from a shared uint8 ([0,255]) or float32 ([0,1]) dtype.
preview_count	integer	no	default 0	Opt in to embedding up to four compact reference/reconstruction PNG thumbnails in the metrics report for portable result documentation.
preview_size	integer	no	default 128	Maximum thumbnail width or height when preview_count is nonzero.
psnr_cap_db	number	no	default 99.0	Requested finite reporting floor for exact reconstructions. The evaluator raises it when necessary so an exact reconstruction always scores above every finite PSNR representable for the evaluated image domain and shape.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

metrics.isac_ofdm

Name: ISAC communication/sensing allocation metrics

Status: implemented

Inputs:

Name	Kind
decision	isac.ofdm_power_allocation.numpy
problem	isac.ofdm_allocation_problem.numpy

Optional inputs:

None.

Outputs:

Name	Kind
report	metrics.report

Parameters:

None.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

metrics.leo_ntn_tracking

Name: LEO-NTN Doppler and handover metrics

Status: implemented

Inputs:

Name	Kind
decision	ntn.future_state_decision.numpy
truth	ntn.future_state_truth.numpy

Optional inputs:

None.

Outputs:

Name	Kind
report	metrics.report

Parameters:

None.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

metrics.localization

Name: Localization metrics

Status: implemented

Inputs:

Name	Kind
estimate	ai_phy.localization_estimate.numpy
problem	ai_phy.localization_problem.numpy, ai_phy.localization_truth.numpy

Optional inputs:

None.

Outputs:

Name	Kind
report	metrics.report

Parameters:

None.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

metrics.modulation_classification

Name: Modulation-recognition metrics

Status: implemented

Inputs:

Name	Kind
prediction	ai_phy.modulation_predictions.numpy
truth	ai_phy.modulation_labels.numpy

Optional inputs:

None.

Outputs:

Name	Kind
confusion_matrix	metrics.confusion_matrix.numpy
report	metrics.report

Parameters:

Name	Type	Required	Default / values	Description
snr_bin_width_db	number	no	default 2.0

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Discrete evaluation metrics are terminal evidence.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy
dataset_capture	numpy
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "json|npz",
  "tensor": "numpy.ndarray"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

metrics.near_field_focusing

Name: Near-field range-angle and focusing metrics

Status: implemented

Inputs:

Name	Kind
estimate	near_field.range_angle_estimate.numpy
problem	near_field.array_observation.numpy
truth	near_field.range_angle_truth.numpy

Optional inputs:

None.

Outputs:

Name	Kind
report	metrics.report

Parameters:

None.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

metrics.nr_ldpc_ofdm_delivery

Name: NR LDPC OFDM delivery metrics (v1)

Status: implemented

Inputs:

Name	Kind
actual_state	channel.ofdm_channel_state.numpy
allocation	channel.power_allocation.numpy
decoded_payload	channel.payload_bits.numpy, channel.bits.numpy
decoder_report	metrics.report
reference_payload	channel.payload_bits.numpy, channel.bits.numpy
tx_symbols	channel.symbols.complex_numpy

Optional inputs:

Name	Kind
rx_symbols	channel.rx_symbols.complex_numpy

Outputs:

Name	Kind
report	metrics.report

Parameters:

None.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

CRC decisions and bit-exact payload delivery are discrete evidence boundaries.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy
dataset_capture	numpy
differentiable_export	None

Equivalence:

{
  "reason": "The v1 score is a deterministic count of CRC-gated payload bits and occupied QPSK data resource elements.",
  "type": "exact"
}

Formats:

{
  "artifact": "json",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

metrics.ofdm_finite_blocklength_allocation

Name: Finite-blocklength delayed-CSI allocation metrics

Status: implemented

Inputs:

Name	Kind
actual_state	channel.ofdm_channel_state.numpy
allocation	channel.power_allocation.numpy
transmitter_csi	channel.ofdm_channel_state.numpy

Optional inputs:

None.

Outputs:

Name	Kind
report	metrics.report

Parameters:

Name	Type	Required	Default / values	Description
blocklength_channel_uses	integer	no	default 128	Short-packet blocklength used by the parallel-channel normal approximation.
include_third_order_term	boolean	no	default True	Include log2(n)/(2n) in the normal approximation.
target_rate_bps_hz	number	no	default 2.0	Fixed payload rate per complex channel use.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

metrics.ofdm_power_allocation

Name: OFDM power-allocation metrics

Status: implemented

Inputs:

Name	Kind
allocation	channel.power_allocation.numpy
state	channel.ofdm_channel_state.numpy

Optional inputs:

Name	Kind
candidate	channel.payload_bits.numpy, channel.bits.numpy
reference	channel.payload_bits.numpy, channel.bits.numpy
symbols	channel.symbols.complex_numpy

Outputs:

Name	Kind
report	metrics.report

Parameters:

Name	Type	Required	Default / values	Description
outage_target_spectral_efficiency_bps_hz	number	no	default 2.0	Target instantaneous Shannon spectral efficiency used for the fading-outage statistic.
transport_block_size_bits	integer	no	default 1024	Payload bits per independently scored transport block for empirical BLER and goodput.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

metrics.resource_allocation

Name: Resource allocation metrics

Status: implemented

Inputs:

Name	Kind
decision	ai_phy.resource_allocation_decision.numpy
problem	ai_phy.resource_allocation_problem.numpy

Optional inputs:

None.

Outputs:

Name	Kind
report	metrics.report

Parameters:

None.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

metrics.retrieval

Name: Image-text retrieval task metrics

Status: implemented

Inputs:

Name	Kind
rankings	retrieval.rankings.json

Optional inputs:

None.

Outputs:

Name	Kind
report	metrics.report

Parameters:

Name	Type	Required	Default / values	Description
k_values	string	no	default 1,5,10

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

metrics.segmentation

Name: Segmentation task metrics

Status: implemented

Inputs:

Name	Kind
candidate	vision.segmentation_mask.numpy
reference	vision.segmentation_mask.numpy

Optional inputs:

None.

Outputs:

Name	Kind
report	metrics.report

Parameters:

None.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

metrics.semantic_state_faithfulness

Name: SemanticState and KB faithfulness metrics

Status: implemented

Inputs:

Name	Kind
candidate	semantic.state.json
kb	foundation.kb.json
reference	semantic.state.json

Optional inputs:

None.

Outputs:

Name	Kind
report	metrics.report

Parameters:

None.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

metrics.text_semantic_similarity

Name: Text lexical similarity metrics

Status: implemented

Inputs:

Name	Kind
candidate	text.batch.json
reference	text.batch.json

Optional inputs:

None.

Outputs:

Name	Kind
report	metrics.report

Parameters:

None.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

metrics.vqa

Name: Visual question answering task metrics

Status: implemented

Inputs:

Name	Kind
candidate	vqa.answers.json, task.predictions.json, task.labels.json
reference	vqa.answers.json, task.labels.json

Optional inputs:

None.

Outputs:

Name	Kind
report	metrics.report

Parameters:

None.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

model

model.aoa_estimator_adapter

Name: AoA estimator adapter

Status: implemented

Inputs:

Name	Kind
problem	ai_phy.aoa_problem.numpy, ai_phy.aoa_observation.numpy

Optional inputs:

None.

Outputs:

Name	Kind
estimate	ai_phy.aoa_estimate.numpy

Parameters:

Name	Type	Required	Default / values	Description
artifact_entrypoint	string	no	default aoa_estimator
artifact_manifest_path	string	no	default ``	Registered schema-v2 trained artifact implementing the ULA AoA ABI.
artifact_package_sha256	string	no	default ``
grid_step_deg	number	no	default 0.25
mode	string	no	default bartlett_reference values music, bartlett_reference, learned_artifact	Select MUSIC, Bartlett, or a returned portable AoA estimator.

Differentiability (legacy trainable_params): framework=torch, gradient=surrogate, trainable_params=True, exportable=True

The adapter accepts a portable learned ULA estimator while MUSIC and Bartlett remain fixed comparison methods.

Training capabilities: built_in_fine_tuning=False, portable_replacement=True

Portable trained-artifact ABI:

{
  "binding_params": {
    "artifact_entrypoint": "aoa_estimator",
    "artifact_manifest_path": "trained_artifact.yaml",
    "mode": "learned_artifact"
  },
  "component_id": "aoa_estimator",
  "component_role": "single_source_ula_aoa_estimator",
  "entrypoint_id": "aoa_estimator",
  "inputs": {
    "snapshots_ri": {
      "dtype": "float32",
      "shape": [
        "batch",
        "antenna",
        "snapshot",
        2
      ]
    }
  },
  "outputs": {
    "angles_deg": {
      "dtype": "float32",
      "shape": [
        "batch"
      ]
    }
  },
  "required_operation_inputs": [
    "problem"
  ]
}

Backends:

Runner	Backends
benchmark_run	numpy, onnxruntime
dataset_capture	numpy, onnxruntime
differentiable_export	torch, sionna

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "npz",
  "tensor": "torch.Tensor"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "music_spatial_spectrum_reference",
    "parameter_bindings": {
      "mode": "music"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "bartlett_spatial_spectrum_reference",
    "parameter_bindings": {
      "mode": "bartlett_reference"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "onnxruntime",
    "implementation": "portable_trained_artifact_runtime",
    "parameter_bindings": {
      "mode": "learned_artifact"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "music_spatial_spectrum_reference",
    "parameter_bindings": {
      "mode": "music"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "bartlett_spatial_spectrum_reference",
    "parameter_bindings": {
      "mode": "bartlett_reference"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "onnxruntime",
    "implementation": "portable_trained_artifact_runtime",
    "parameter_bindings": {
      "mode": "learned_artifact"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "trainable_aoa_estimator_endpoint",
    "runner": "differentiable_export",
    "status": "implemented"
  },
  {
    "backend": "sionna",
    "implementation": "sionna_rt_aoa_training_endpoint",
    "runner": "differentiable_export",
    "status": "planned"
  }
]

model.beamforming_adapter

Name: Beamforming/precoding adapter

Status: implemented

Inputs:

Name	Kind
problem	ai_phy.beamforming_problem.numpy

Optional inputs:

None.

Outputs:

Name	Kind
decision	ai_phy.beamforming_decision.numpy

Parameters:

Name	Type	Required	Default / values	Description
artifact_entrypoint	string	no	default beam_policy
artifact_manifest_path	string	no	default ``	Registered schema-v2 trained artifact implementing the beam-policy ABI.
artifact_package_sha256	string	no	default ``
mode	string	no	default codebook_sweep_reference values mrt, codebook_sweep_reference, learned_artifact	Select a fixed reference or a returned portable beam-policy artifact.

Differentiability (legacy trainable_params): framework=torch, gradient=surrogate, trainable_params=True, exportable=True

The adapter accepts a portable learned beam policy while MRT and exhaustive DFT-codebook selection remain fixed comparison methods.

Training capabilities: built_in_fine_tuning=False, portable_replacement=True

Portable trained-artifact ABI:

{
  "binding_params": {
    "artifact_entrypoint": "beam_policy",
    "artifact_manifest_path": "trained_artifact.yaml",
    "mode": "learned_artifact"
  },
  "component_id": "beam_policy",
  "component_role": "single_user_miso_beam_policy",
  "entrypoint_id": "beam_policy",
  "inputs": {
    "channels_ri": {
      "dtype": "float32",
      "shape": [
        "batch",
        "tx_antenna",
        2
      ]
    }
  },
  "outputs": {
    "weights_ri": {
      "dtype": "float32",
      "shape": [
        "batch",
        "tx_antenna",
        2
      ]
    }
  },
  "required_operation_inputs": [
    "problem"
  ]
}

Backends:

Runner	Backends
benchmark_run	numpy, onnxruntime
dataset_capture	numpy, onnxruntime
differentiable_export	torch, sionna

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "maximum_ratio_transmission",
    "parameter_bindings": {
      "mode": "mrt"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "codebook_sweep_reference",
    "parameter_bindings": {
      "mode": "codebook_sweep_reference"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "onnxruntime",
    "implementation": "portable_trained_artifact_runtime",
    "parameter_bindings": {
      "mode": "learned_artifact"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "maximum_ratio_transmission",
    "parameter_bindings": {
      "mode": "mrt"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "codebook_sweep_reference",
    "parameter_bindings": {
      "mode": "codebook_sweep_reference"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "onnxruntime",
    "implementation": "portable_trained_artifact_runtime",
    "parameter_bindings": {
      "mode": "learned_artifact"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "trainable_beam_policy_endpoint",
    "runner": "differentiable_export",
    "status": "implemented"
  },
  {
    "backend": "sionna",
    "implementation": "sionna_precoder_training_endpoint",
    "runner": "differentiable_export",
    "status": "planned"
  }
]

model.causal_csi_power_allocator

Name: Causal-history symbol power allocator

Status: implemented

Inputs:

Name	Kind
symbols	channel.symbols.complex_numpy, channel.rx_symbols.complex_numpy

Optional inputs:

Name	Kind
bits	channel.payload_bits.numpy, channel.coded_bits.numpy, channel.bits.numpy
channel_state	channel.ofdm_channel_state.numpy

Outputs:

Name	Kind
allocation	channel.power_allocation.numpy
symbols	channel.symbols.complex_numpy

Parameters:

Name	Type	Required	Default / values	Description
allocation_contrast	number	no	default 0.6
allocation_lower_power_ratio	number	no	default 0.1	Minimum per-subcarrier power relative to the mean budget for bounded causal-AR water filling.
allocation_upper_power_ratio	number	no	default 1.9	Maximum per-subcarrier power relative to the mean budget for bounded causal-AR water filling.
artifact_entrypoint	string	no	default power_policy	Entrypoint in the registered artifact manifest used for allocator inference.
artifact_manifest_path	string	no	default ``	Registered trained-artifact manifest implementing the architecture-neutral allocator slot ABI.
artifact_package_sha256	string	no	default ``	Registry-independent digest of the complete trained-artifact package bound into the execution plan.
bit_loading_bpsk_min_snr_db	number	no	default 6.0
bit_loading_max_bits_per_symbol	integer	no	default 6 values 1, 2, 4, 6
bit_loading_qam16_min_snr_db	number	no	default 17.0
bit_loading_qam64_min_snr_db	number	no	default 23.0
bit_loading_qpsk_min_snr_db	number	no	default 10.0
budget_mode	string	no	default fixed_average values fixed_average, variable_average
checkpoint_format	string	no	default noema_csi_power_deepset_npz_v1 values noema_csi_power_deepset_npz_v1
checkpoint_max_bytes	integer	no	default 67108864
checkpoint_path	string	no	default ``	Frozen safe-NPZ CSI allocator checkpoint used by policy=learned_checkpoint.
checkpoint_sha256	string	no	default ``	Required lowercase SHA-256 of the frozen allocator checkpoint.
checkpoint_strict	boolean	no	default True
csi_gain_shrinkage	number	no	default 0.6	For uncertainty-shrunk water filling, blend observed gains toward their per-state mean before allocating power.
csi_prediction_gain_confidence	number	no	default 0.4	Blend complex-AR predicted gains toward their per-state frequency mean before water filling.
csi_prediction_horizon_ofdm_symbols	integer	no	default 0	For complex-AR prediction, forecast this many OFDM symbols past the newest causal CSI snapshot. Zero uses the CSI artifact’s declared feedback delay.
eps	number	no	default 1e-12
granularity	string	no	default global values global, per_symbol, per_subcarrier, per_stream
label	string	no	default tx_power_allocate
max_power	number	no	default 2.0
midpoint_snr_db	number	no	default 12.0
min_power	number	no	default 0.25
model_batch_size	integer	no	default 1024	Maximum number of independent channel-state examples evaluated in one learned-model inference call. Chunking preserves the complete experiment batch and output order.
policy	string	no	default snr_sigmoid values fixed, snr_sigmoid, water_filling, observed_csi_water_filling, robust_csi_water_filling, causal_ar_water_filling, causal_ar_box_water_filling, learned_checkpoint, learned_artifact	Power-allocation method. Model selection is shown only for learned policies.
power_unit	string	no	default normalized values normalized, mW
slope_db	number	no	default 4.0
snr_db	number	no	default 12.0	Fallback reference SNR used only by the SNR-adaptive sigmoid policy when no explicit channel state is connected.
stream_count	integer	no	default 1
subcarrier_count	integer	no	default 64
target_power	number	no	default 1.0	Normalized average complex-symbol power budget per resource element; the per-state sum constraint is this value times the subcarrier count.
transport_mode	string	no	default fixed_modulation values fixed_modulation, allocation_aware_bit_loading	Optionally convert the coded-bit stream into a CSI/allocation-aware OFDM modulation map.

Differentiability (legacy trainable_params): framework=torch, gradient=full, trainable_params=True, exportable=True

Continuous-symbol power allocation is differentiable. Noema can benchmark a frozen, CSI-conditioned artifact behind an architecture-neutral runtime ABI while model training remains outside the benchmark executor.

Training capabilities: built_in_fine_tuning=False, portable_replacement=True

Portable trained-artifact ABI:

{
  "binding_params": {
    "artifact_entrypoint": "power_policy",
    "artifact_manifest_path": "trained_artifact.yaml",
    "budget_mode": "fixed_average",
    "granularity": "per_subcarrier",
    "policy": "learned_artifact"
  },
  "component_id": "policy",
  "component_role": "power_policy",
  "constraint_adapter": "noema_exact_simplex_projection_v1",
  "entrypoint_id": "power_policy",
  "inputs": {
    "average_power_budget": {
      "dtype": "float32",
      "shape": [
        "batch",
        1
      ]
    },
    "csi_history": {
      "dtype": "float32",
      "shape": [
        "batch",
        "history",
        "subcarrier",
        2
      ]
    },
    "noise_variance": {
      "dtype": "float32",
      "shape": [
        "batch",
        1
      ]
    }
  },
  "outputs": {
    "allocation_scores": {
      "dtype": "float32",
      "shape": [
        "batch",
        "subcarrier"
      ]
    }
  },
  "required_operation_inputs": [
    "channel_state"
  ]
}

Backends:

Runner	Backends
benchmark_run	numpy, onnxruntime
dataset_capture	numpy, onnxruntime
differentiable_export	torch

Equivalence:

{
  "reason": "Reference NumPy and Torch policies should select the same group power targets and scaled symbols for fixed parameters.",
  "tolerance": {
    "atol": 1e-06,
    "rtol": 1e-05
  },
  "type": "numerical"
}

Formats:

{
  "artifact": "npz",
  "checkpoint": "noema_csi_power_deepset_npz_v1",
  "tensor": "torch.Tensor"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "fixed_power_allocator",
    "parameter_bindings": {
      "policy": "fixed"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "snr_sigmoid_power_allocator",
    "parameter_bindings": {
      "policy": "snr_sigmoid"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "closed_form_water_filling_allocator",
    "parameter_bindings": {
      "policy": "water_filling"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "water_filling_on_observed_csi",
    "parameter_bindings": {
      "policy": "observed_csi_water_filling"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "correlation_shrunk_water_filling",
    "parameter_bindings": {
      "policy": "robust_csi_water_filling"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "causal_complex_ar_prediction_plus_water_filling",
    "parameter_bindings": {
      "policy": "causal_ar_water_filling"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "causal_complex_ar_prediction_plus_box_constrained_water_filling",
    "parameter_bindings": {
      "policy": "causal_ar_box_water_filling"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "onnxruntime",
    "implementation": "portable_trained_artifact_policy_plus_noema_projection",
    "parameter_bindings": {
      "policy": "learned_artifact"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "fixed_power_allocator",
    "parameter_bindings": {
      "policy": "fixed"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "snr_sigmoid_power_allocator",
    "parameter_bindings": {
      "policy": "snr_sigmoid"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "closed_form_water_filling_allocator",
    "parameter_bindings": {
      "policy": "water_filling"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "water_filling_on_observed_csi",
    "parameter_bindings": {
      "policy": "observed_csi_water_filling"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "correlation_shrunk_water_filling",
    "parameter_bindings": {
      "policy": "robust_csi_water_filling"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "causal_complex_ar_prediction_plus_water_filling",
    "parameter_bindings": {
      "policy": "causal_ar_water_filling"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "causal_complex_ar_prediction_plus_box_constrained_water_filling",
    "parameter_bindings": {
      "policy": "causal_ar_box_water_filling"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "onnxruntime",
    "implementation": "portable_trained_artifact_policy_plus_noema_projection",
    "parameter_bindings": {
      "policy": "learned_artifact"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "trainable_budgeted_symbol_allocator",
    "runner": "differentiable_export",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "safe_npz_csi_deepset_checkpoint",
    "parameter_bindings": {
      "policy": "learned_checkpoint"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "safe_npz_csi_deepset_checkpoint",
    "parameter_bindings": {
      "policy": "learned_checkpoint"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

model.channel_estimator_adapter

Name: Channel estimator adapter

Status: implemented

Inputs:

Name	Kind
problem	ai_phy.channel_estimation_problem.numpy, ai_phy.channel_estimation_observation.numpy

Optional inputs:

None.

Outputs:

Name	Kind
estimate	ai_phy.channel_estimate.numpy

Parameters:

Name	Type	Required	Default / values	Description
artifact_entrypoint	string	no	default channel_estimator
artifact_manifest_path	string	no	default ``	Registered schema-v2 trained artifact implementing the MIMO-OFDM channel-estimator ABI.
artifact_package_sha256	string	no	default ``
lmmse_assumed_tap_count	integer	no	default 4	Fixed exponential-PDP prior used by the practical LMMSE baseline. It is not adapted from hidden channel truth.
mode	string	no	default linear_mmse_reference values least_squares, linear_mmse_reference, learned_artifact	Runnable estimator. The learned mode uses a returned portable artifact; the other modes are fixed baselines.

Differentiability (legacy trainable_params): framework=torch, gradient=surrogate, trainable_params=True, exportable=True

The estimator slot accepts a portable learned artifact while classical LS and exponential-PDP LMMSE modes remain reproducible baselines.

Training capabilities: built_in_fine_tuning=False, portable_replacement=True

Portable trained-artifact ABI:

{
  "binding_params": {
    "artifact_entrypoint": "channel_estimator",
    "artifact_manifest_path": "trained_artifact.yaml",
    "mode": "learned_artifact"
  },
  "component_id": "estimator",
  "component_role": "mimo_ofdm_channel_estimator",
  "entrypoint_id": "channel_estimator",
  "inputs": {
    "ls_estimate_ri": {
      "dtype": "float32",
      "shape": [
        "batch",
        "rx_antenna",
        "tx_antenna",
        "subcarrier",
        2
      ]
    },
    "noise_variance": {
      "dtype": "float32",
      "shape": [
        "batch",
        1
      ]
    },
    "pilot_ls_ri": {
      "dtype": "float32",
      "shape": [
        "batch",
        "rx_antenna",
        "tx_antenna",
        "subcarrier",
        2
      ]
    },
    "pilot_mask": {
      "dtype": "float32",
      "shape": [
        "batch",
        "tx_antenna",
        "subcarrier"
      ]
    }
  },
  "outputs": {
    "h_hat_ri": {
      "dtype": "float32",
      "shape": [
        "batch",
        "rx_antenna",
        "tx_antenna",
        "subcarrier",
        2
      ]
    }
  },
  "required_operation_inputs": [
    "problem"
  ]
}

Backends:

Runner	Backends
benchmark_run	numpy, onnxruntime
dataset_capture	numpy, onnxruntime
differentiable_export	torch, sionna

Equivalence:

{
  "reason": "Adapter implementations are compared by NMSE/task metrics under the declared benchmark protocol.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "npz",
  "tensor": "torch.Tensor"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "least_squares",
    "parameter_bindings": {
      "mode": "least_squares"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "linear_mmse_reference",
    "parameter_bindings": {
      "mode": "linear_mmse_reference"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "onnxruntime",
    "implementation": "portable_trained_artifact_runtime",
    "parameter_bindings": {
      "mode": "learned_artifact"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "least_squares",
    "parameter_bindings": {
      "mode": "least_squares"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "linear_mmse_reference",
    "parameter_bindings": {
      "mode": "linear_mmse_reference"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "onnxruntime",
    "implementation": "portable_trained_artifact_runtime",
    "parameter_bindings": {
      "mode": "learned_artifact"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "trainable_channel_estimator_endpoint",
    "runner": "differentiable_export",
    "status": "implemented"
  },
  {
    "backend": "sionna",
    "implementation": "sionna_receiver_training_endpoint",
    "runner": "differentiable_export",
    "status": "implemented"
  }
]

model.compressai_analysis_encode

Name: CompressAI PyTorch analysis transform to latents

Status: implemented

Inputs:

Name	Kind
images	image.batch.numpy

Optional inputs:

None.

Outputs:

Name	Kind
latents	semantic.latents.numpy

Parameters:

Name	Type	Required	Default / values	Description
device	string	no	default cpu
metric	string	no	default mse values mse, ms-ssim
model	string	no	default bmshj2018_hyperprior values bmshj2018_factorized, bmshj2018_factorized_relu, bmshj2018_hyperprior, bmshj2018_hyperprior_vbr, mbt2018_mean, mbt2018_mean_vbr, mbt2018, mbt2018_vbr, cheng2020_anchor, cheng2020_attn
pad_to_multiple	integer	no	default 64
pretrained	boolean	no	default True
quality	integer	no	default 3
quantize_latents	boolean	no	default False

Differentiability (legacy trainable_params): framework=torch, gradient=full, trainable_params=False, exportable=True

The analysis transform is a PyTorch module; Noema benchmark execution runs pretrained eval parameters.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

model.compressai_aoti_decode

Name: CompressAI AOT Inductor synthesis transform to image batch

Status: implemented

Inputs:

Name	Kind
latents	semantic.latents.numpy
model	model.aot_inductor.bundle

Optional inputs:

None.

Outputs:

Name	Kind
images	image.batch.numpy

Parameters:

Name	Type	Required	Default / values	Description
device	string	no	default cpu
on_error	string	no	default fail values fail, zeros

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

model.compressai_aoti_encode

Name: CompressAI AOT Inductor analysis transform to latents

Status: implemented

Inputs:

Name	Kind
images	image.batch.numpy
model	model.aot_inductor.bundle

Optional inputs:

None.

Outputs:

Name	Kind
latents	semantic.latents.numpy

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
device	string	no	default cpu
pad_to_multiple	integer	no	default 64
quantize_latents	boolean	no	default False

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

model.compressai_aoti_entropy_decode

Name: CompressAI AOT Inductor payload decode bits to latents

Status: implemented

Inputs:

Name	Kind
bits	channel.payload_bits.numpy, channel.bits.numpy
model	model.aot_inductor.bundle

Optional inputs:

None.

Outputs:

Name	Kind
latents	semantic.latents.numpy

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
device	string	no	default cpu
on_error	string	no	default fail values fail, zeros

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

Entropy/payload info:

{
  "algorithm": "range coding / ANS backend",
  "coder": "CompressAI entropy bottleneck + GaussianConditional",
  "implementation": "AOT Inductor hyperprior transforms + CompressAI entropy backend",
  "language": "AOT Inductor/PyTorch + C++ extension + Python",
  "note": "The neural transforms may run in PyTorch, AOT Inductor, ONNX Runtime, or OpenVINO; entropy coding remains the CompressAI bitstream backend."
}

model.compressai_aoti_entropy_encode

Name: CompressAI AOT Inductor payload encode latents to bits

Status: implemented

Inputs:

Name	Kind
latents	semantic.latents.numpy
model	model.aot_inductor.bundle

Optional inputs:

None.

Outputs:

Name	Kind
bits	channel.payload_bits.numpy

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
device	string	no	default cpu

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

Entropy/payload info:

{
  "algorithm": "range coding / ANS backend",
  "coder": "CompressAI entropy bottleneck + GaussianConditional",
  "implementation": "AOT Inductor hyperprior transforms + CompressAI entropy backend",
  "language": "AOT Inductor/PyTorch + C++ extension + Python",
  "note": "The neural transforms may run in PyTorch, AOT Inductor, ONNX Runtime, or OpenVINO; entropy coding remains the CompressAI bitstream backend."
}

model.compressai_decode

Name: CompressAI model-zoo payload bits decoder to image batch

Status: implemented

Inputs:

Name	Kind
bits	channel.payload_bits.numpy, channel.bits.numpy

Optional inputs:

None.

Outputs:

Name	Kind
images	image.batch.numpy

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
device	string	no	default cpu
on_error	string	no	default fail values fail, zeros, gray_image, erasure, report_outage

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

Entropy/payload info:

{
  "algorithm": "range coding / ANS backend",
  "coder": "CompressAI entropy bottleneck + GaussianConditional",
  "implementation": "CompressAI entropy backend",
  "language": "C++ extension + Python/PyTorch",
  "note": "The neural transforms may run in PyTorch, AOT Inductor, ONNX Runtime, or OpenVINO; entropy coding remains the CompressAI bitstream backend."
}

model.compressai_encode

Name: CompressAI model-zoo image encoder to payload bits

Status: implemented

Inputs:

Name	Kind
images	image.batch.numpy

Optional inputs:

None.

Outputs:

Name	Kind
bits	channel.payload_bits.numpy

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
device	string	no	default cpu
expected_model_state_sha256	string	no	default ``	Optional publication pin for the fully materialized CompressAI state after model.update().
metric	string	no	default mse values mse, ms-ssim
model	string	no	default bmshj2018_hyperprior values bmshj2018_factorized, bmshj2018_factorized_relu, bmshj2018_hyperprior, bmshj2018_hyperprior_vbr, mbt2018_mean, mbt2018_mean_vbr, mbt2018, mbt2018_vbr, cheng2020_anchor, cheng2020_attn
pad_to_multiple	integer	no	default 64
pretrained	boolean	no	default True
quality	integer	no	default 3
vbr_scale_index	integer	no	default 1	Only used by CompressAI *_vbr models. Selects the variable-rate scale index s, where larger values usually mean higher rate and quality.
vbr_stage	integer	no	default 2	Only used by CompressAI *_vbr models. Stage 2 uses the variable-rate path; stage 1 behaves like the base model path.
wire_format	string	no	default safe_json_base64 values safe_json_base64, compact_binary_v1	Artifact-safe JSON/base64 remains the compatibility default; compact_binary_v1 is the bounded communication wire format for transmitted-resource studies.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

Entropy/payload info:

{
  "algorithm": "range coding / ANS backend",
  "coder": "CompressAI entropy bottleneck + GaussianConditional",
  "implementation": "CompressAI entropy backend",
  "language": "C++ extension + Python/PyTorch",
  "note": "The neural transforms may run in PyTorch, AOT Inductor, ONNX Runtime, or OpenVINO; entropy coding remains the CompressAI bitstream backend."
}

model.compressai_entropy_decode

Name: CompressAI payload decode bits to latents

Status: implemented

Inputs:

Name	Kind
bits	channel.payload_bits.numpy, channel.bits.numpy

Optional inputs:

None.

Outputs:

Name	Kind
latents	semantic.latents.numpy

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
device	string	no	default cpu
on_error	string	no	default fail values fail, zeros

Differentiability (legacy trainable_params): framework=torch, gradient=stop, trainable_params=False, exportable=False

Entropy decoding and bitstream parsing are discrete payload-coding steps.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

Entropy/payload info:

{
  "algorithm": "range coding / ANS backend",
  "coder": "CompressAI entropy bottleneck + GaussianConditional",
  "implementation": "CompressAI entropy backend",
  "language": "C++ extension + Python/PyTorch",
  "note": "The neural transforms may run in PyTorch, AOT Inductor, ONNX Runtime, or OpenVINO; entropy coding remains the CompressAI bitstream backend."
}

model.compressai_entropy_encode

Name: CompressAI payload encode latents to bits

Status: implemented

Inputs:

Name	Kind
latents	semantic.latents.numpy

Optional inputs:

None.

Outputs:

Name	Kind
bits	channel.payload_bits.numpy

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
device	string	no	default cpu
metric	string	no	default mse values mse, ms-ssim
model	string	no	default bmshj2018_hyperprior values bmshj2018_factorized, bmshj2018_factorized_relu, bmshj2018_hyperprior, bmshj2018_hyperprior_vbr, mbt2018_mean, mbt2018_mean_vbr, mbt2018, mbt2018_vbr, cheng2020_anchor, cheng2020_attn
pretrained	boolean	no	default True
quality	integer	no	default 3

Differentiability (legacy trainable_params): framework=torch, gradient=stop, trainable_params=False, exportable=False

Entropy coding and bitstream serialization are discrete payload-coding steps.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

Entropy/payload info:

{
  "algorithm": "range coding / ANS backend",
  "coder": "CompressAI entropy bottleneck + GaussianConditional",
  "implementation": "CompressAI entropy backend",
  "language": "C++ extension + Python/PyTorch",
  "note": "The neural transforms may run in PyTorch, AOT Inductor, ONNX Runtime, or OpenVINO; entropy coding remains the CompressAI bitstream backend."
}

model.compressai_export_aoti

Name: Compile CompressAI PyTorch transforms with AOT Inductor

Status: implemented

Inputs:

Name	Kind
images	image.batch.numpy

Optional inputs:

None.

Outputs:

Name	Kind
model	model.aot_inductor.bundle

Parameters:

Name	Type	Required	Default / values	Description
device	string	no	default cpu
dynamic_shapes	boolean	no	default False	Advanced. Keep off for reproducible selected-shape AOTI packages.
export_height	integer	no	default 0	Advanced. 0 compiles for the selected data shape after padding.
export_width	integer	no	default 0	Advanced. 0 compiles for the selected data shape after padding.
metric	string	no	default mse values mse, ms-ssim
model	string	no	default bmshj2018_hyperprior values bmshj2018_factorized, bmshj2018_factorized_relu, bmshj2018_hyperprior, bmshj2018_hyperprior_vbr, mbt2018_mean, mbt2018_mean_vbr, mbt2018, mbt2018_vbr, cheng2020_anchor, cheng2020_attn
pad_to_multiple	integer	no	default 64
pretrained	boolean	no	default True
quality	integer	no	default 3
validate_export	boolean	no	default True
vbr_scale_index	integer	no	default 1
vbr_stage	integer	no	default 2

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

model.compressai_export_onnx

Name: Convert CompressAI PyTorch transforms to ONNX

Status: implemented

Inputs:

Name	Kind
images	image.batch.numpy

Optional inputs:

None.

Outputs:

Name	Kind
model	model.onnx.bundle

Parameters:

Name	Type	Required	Default / values	Description
device	string	no	default cpu
export_height	integer	no	default 0	Advanced. 0 exports for the selected data shape after padding.
export_width	integer	no	default 0	Advanced. 0 exports for the selected data shape after padding.
metric	string	no	default mse values mse, ms-ssim
model	string	no	default bmshj2018_hyperprior values bmshj2018_factorized, bmshj2018_factorized_relu, bmshj2018_hyperprior, bmshj2018_hyperprior_vbr, mbt2018_mean, mbt2018_mean_vbr, mbt2018, mbt2018_vbr, cheng2020_anchor, cheng2020_attn
opset	integer	no	default 17
pad_to_multiple	integer	no	default 64
pretrained	boolean	no	default True
quality	integer	no	default 3
validate_export	boolean	no	default True
vbr_scale_index	integer	no	default 1
vbr_stage	integer	no	default 2

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

model.compressai_onnx_cpp_decode

Name: CompressAI ONNX Runtime C++ API latent decoder to image batch

Status: implemented

Inputs:

Name	Kind
latents	semantic.latents.numpy
model	model.onnx.bundle

Optional inputs:

None.

Outputs:

Name	Kind
images	image.batch.numpy

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
intra_op_num_threads	integer	no	default 0
library_path	string	no	default ``
on_error	string	no	default fail values fail, zeros
provider	string	no	default CPUExecutionProvider

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

model.compressai_onnx_cpp_encode

Name: CompressAI ONNX Runtime C++ API analysis transform to latents

Status: implemented

Inputs:

Name	Kind
images	image.batch.numpy
model	model.onnx.bundle

Optional inputs:

None.

Outputs:

Name	Kind
latents	semantic.latents.numpy

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
intra_op_num_threads	integer	no	default 0
library_path	string	no	default ``
pad_to_multiple	integer	no	default 64
provider	string	no	default CPUExecutionProvider
quantize_latents	boolean	no	default False

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

model.compressai_onnx_cpp_entropy_decode

Name: CompressAI ONNX Runtime C++ API payload decode bits to latents

Status: implemented

Inputs:

Name	Kind
bits	channel.payload_bits.numpy, channel.bits.numpy
model	model.onnx.bundle

Optional inputs:

None.

Outputs:

Name	Kind
latents	semantic.latents.numpy

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
device	string	no	default cpu
intra_op_num_threads	integer	no	default 0
library_path	string	no	default ``
on_error	string	no	default fail values fail, zeros
provider	string	no	default CPUExecutionProvider

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

Entropy/payload info:

{
  "algorithm": "range coding / ANS backend",
  "coder": "CompressAI entropy bottleneck + GaussianConditional",
  "implementation": "ONNX Runtime hyperprior transforms + CompressAI entropy backend",
  "language": "ONNX Runtime + C++ extension + Python/PyTorch",
  "note": "The neural transforms may run in PyTorch, AOT Inductor, ONNX Runtime, or OpenVINO; entropy coding remains the CompressAI bitstream backend."
}

model.compressai_onnx_cpp_entropy_encode

Name: CompressAI ONNX Runtime C++ API payload encode latents to bits

Status: implemented

Inputs:

Name	Kind
latents	semantic.latents.numpy
model	model.onnx.bundle

Optional inputs:

None.

Outputs:

Name	Kind
bits	channel.payload_bits.numpy

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
device	string	no	default cpu
intra_op_num_threads	integer	no	default 0
library_path	string	no	default ``
provider	string	no	default CPUExecutionProvider

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

Entropy/payload info:

{
  "algorithm": "range coding / ANS backend",
  "coder": "CompressAI entropy bottleneck + GaussianConditional",
  "implementation": "ONNX Runtime hyperprior transforms + CompressAI entropy backend",
  "language": "ONNX Runtime + C++ extension + Python/PyTorch",
  "note": "The neural transforms may run in PyTorch, AOT Inductor, ONNX Runtime, or OpenVINO; entropy coding remains the CompressAI bitstream backend."
}

model.compressai_onnx_decode

Name: CompressAI ONNX Runtime quantized latent decoder to image batch

Status: implemented

Inputs:

Name	Kind
latents	semantic.latents.numpy
model	model.onnx.bundle

Optional inputs:

None.

Outputs:

Name	Kind
images	image.batch.numpy

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
on_error	string	no	default fail values fail, zeros
provider	string	no	default CPUExecutionProvider

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

model.compressai_onnx_encode

Name: CompressAI ONNX Runtime analysis transform to latents

Status: implemented

Inputs:

Name	Kind
images	image.batch.numpy
model	model.onnx.bundle

Optional inputs:

None.

Outputs:

Name	Kind
latents	semantic.latents.numpy

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
pad_to_multiple	integer	no	default 64
provider	string	no	default CPUExecutionProvider
quantize_latents	boolean	no	default False

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

model.compressai_onnx_entropy_decode

Name: CompressAI ONNX payload decode bits to latents

Status: implemented

Inputs:

Name	Kind
bits	channel.payload_bits.numpy, channel.bits.numpy
model	model.onnx.bundle

Optional inputs:

None.

Outputs:

Name	Kind
latents	semantic.latents.numpy

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
device	string	no	default cpu
on_error	string	no	default fail values fail, zeros
provider	string	no	default CPUExecutionProvider

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

Entropy/payload info:

{
  "algorithm": "range coding / ANS backend",
  "coder": "CompressAI entropy bottleneck + GaussianConditional",
  "implementation": "ONNX Runtime hyperprior transforms + CompressAI entropy backend",
  "language": "ONNX Runtime + C++ extension + Python/PyTorch",
  "note": "The neural transforms may run in PyTorch, AOT Inductor, ONNX Runtime, or OpenVINO; entropy coding remains the CompressAI bitstream backend."
}

model.compressai_onnx_entropy_encode

Name: CompressAI ONNX payload encode latents to bits

Status: implemented

Inputs:

Name	Kind
latents	semantic.latents.numpy
model	model.onnx.bundle

Optional inputs:

None.

Outputs:

Name	Kind
bits	channel.payload_bits.numpy

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
device	string	no	default cpu
provider	string	no	default CPUExecutionProvider

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

Entropy/payload info:

{
  "algorithm": "range coding / ANS backend",
  "coder": "CompressAI entropy bottleneck + GaussianConditional",
  "implementation": "ONNX Runtime hyperprior transforms + CompressAI entropy backend",
  "language": "ONNX Runtime + C++ extension + Python/PyTorch",
  "note": "The neural transforms may run in PyTorch, AOT Inductor, ONNX Runtime, or OpenVINO; entropy coding remains the CompressAI bitstream backend."
}

model.compressai_openvino_decode

Name: CompressAI OpenVINO synthesis transform to image batch

Status: implemented

Inputs:

Name	Kind
latents	semantic.latents.numpy
model	model.onnx.bundle

Optional inputs:

None.

Outputs:

Name	Kind
images	image.batch.numpy

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
device	string	no	default CPU
on_error	string	no	default fail values fail, zeros

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

model.compressai_openvino_encode

Name: CompressAI OpenVINO analysis transform to latents

Status: implemented

Inputs:

Name	Kind
images	image.batch.numpy
model	model.onnx.bundle

Optional inputs:

None.

Outputs:

Name	Kind
latents	semantic.latents.numpy

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
device	string	no	default CPU
pad_to_multiple	integer	no	default 64
quantize_latents	boolean	no	default False

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

model.compressai_openvino_entropy_decode

Name: CompressAI OpenVINO payload decode bits to latents

Status: implemented

Inputs:

Name	Kind
bits	channel.payload_bits.numpy, channel.bits.numpy
model	model.onnx.bundle

Optional inputs:

None.

Outputs:

Name	Kind
latents	semantic.latents.numpy

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
device	string	no	default CPU
on_error	string	no	default fail values fail, zeros
torch_device	string	no	default cpu

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

Entropy/payload info:

{
  "algorithm": "range coding / ANS backend",
  "coder": "CompressAI entropy bottleneck + GaussianConditional",
  "implementation": "OpenVINO hyperprior transforms + CompressAI entropy backend",
  "language": "OpenVINO + C++ extension + Python/PyTorch",
  "note": "The neural transforms may run in PyTorch, AOT Inductor, ONNX Runtime, or OpenVINO; entropy coding remains the CompressAI bitstream backend."
}

model.compressai_openvino_entropy_encode

Name: CompressAI OpenVINO payload encode latents to bits

Status: implemented

Inputs:

Name	Kind
latents	semantic.latents.numpy
model	model.onnx.bundle

Optional inputs:

None.

Outputs:

Name	Kind
bits	channel.payload_bits.numpy

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
device	string	no	default CPU
torch_device	string	no	default cpu

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

Entropy/payload info:

{
  "algorithm": "range coding / ANS backend",
  "coder": "CompressAI entropy bottleneck + GaussianConditional",
  "implementation": "OpenVINO hyperprior transforms + CompressAI entropy backend",
  "language": "OpenVINO + C++ extension + Python/PyTorch",
  "note": "The neural transforms may run in PyTorch, AOT Inductor, ONNX Runtime, or OpenVINO; entropy coding remains the CompressAI bitstream backend."
}

model.compressai_synthesis_decode

Name: CompressAI PyTorch synthesis transform to image batch

Status: implemented

Inputs:

Name	Kind
latents	semantic.latents.numpy

Optional inputs:

None.

Outputs:

Name	Kind
images	image.batch.numpy

Parameters:

Name	Type	Required	Default / values	Description
device	string	no	default cpu
on_error	string	no	default fail values fail, zeros

Differentiability (legacy trainable_params): framework=torch, gradient=full, trainable_params=False, exportable=True

The synthesis transform is a PyTorch module; Noema benchmark execution runs pretrained eval parameters.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

model.csi_feedback_decoder

Name: CSI feedback decoder / returned-artifact slot

Status: implemented

Inputs:

Name	Kind
received_code	channel.csi_feedback_received.numpy, channel.csi_feedback_code.numpy

Optional inputs:

None.

Outputs:

Name	Kind
reconstruction	channel.miso_ofdm_csi_reconstruction.numpy

Parameters:

Name	Type	Required	Default / values	Description
artifact_entrypoint	string	no	default decoder
artifact_manifest_path	string	no	default ``	Registered schema-v2 trained-artifact manifest for this paired CSI codec.
artifact_package_sha256	string	no	default ``
feedback_dimension	integer	no	default 64	Number of real-valued feedback latents per CSI realization.
runtime	string	no	default training_interface values training_interface, truncated_angular_delay, identity, learned_artifact	Use the architecture-neutral training slot, a matched-budget classical angular-delay truncation, the full-CSI identity upper bound, or a returned artifact.

Differentiability (legacy trainable_params): framework=torch, gradient=full, trainable_params=True, exportable=True

The recipe declares the feedback-to-CSI interface; external training supplies its architecture.

Training capabilities: built_in_fine_tuning=False, portable_replacement=True

Portable trained-artifact ABI:

{
  "binding_params": {
    "artifact_entrypoint": "decoder",
    "artifact_manifest_path": "trained_artifact.yaml",
    "runtime": "learned_artifact"
  },
  "component_id": "decoder",
  "component_role": "csi_feedback_decoder",
  "entrypoint_id": "decoder",
  "inputs": {
    "feedback_code": {
      "dtype": "float32",
      "shape": [
        "batch",
        "feedback_dimension"
      ]
    }
  },
  "outputs": {
    "csi_hat_ri": {
      "dtype": "float32",
      "shape": [
        "batch",
        2,
        "tx_antenna",
        "subcarrier"
      ]
    }
  },
  "required_operation_inputs": [
    "received_code"
  ]
}

Backends:

Runner	Backends
benchmark_run	numpy, onnxruntime
dataset_capture	numpy, onnxruntime
differentiable_export	torch

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "npz",
  "checkpoint": "onnx",
  "tensor": "float32"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "identity_full_csi_decoder",
    "parameter_bindings": {
      "runtime": "identity"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "truncated_angular_delay_decoder",
    "parameter_bindings": {
      "runtime": "truncated_angular_delay"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "onnxruntime",
    "implementation": "portable_trained_artifact_decoder",
    "parameter_bindings": {
      "runtime": "learned_artifact"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "identity_full_csi_decoder",
    "parameter_bindings": {
      "runtime": "identity"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "truncated_angular_delay_decoder",
    "parameter_bindings": {
      "runtime": "truncated_angular_delay"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "onnxruntime",
    "implementation": "portable_trained_artifact_decoder",
    "parameter_bindings": {
      "runtime": "learned_artifact"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "trainable_csi_feedback_decoder_slot",
    "parameter_bindings": {
      "runtime": "training_interface"
    },
    "runner": "differentiable_export",
    "status": "implemented"
  }
]

model.csi_feedback_encoder

Name: CSI feedback encoder / returned-artifact slot

Status: implemented

Inputs:

Name	Kind
csi	channel.miso_ofdm_csi.numpy

Optional inputs:

None.

Outputs:

Name	Kind
feedback_code	channel.csi_feedback_code.numpy

Parameters:

Name	Type	Required	Default / values	Description
artifact_entrypoint	string	no	default encoder
artifact_manifest_path	string	no	default ``	Registered schema-v2 trained-artifact manifest for this paired CSI codec.
artifact_package_sha256	string	no	default ``
feedback_dimension	integer	no	default 64	Number of real-valued feedback latents per CSI realization.
runtime	string	no	default training_interface values training_interface, truncated_angular_delay, identity, learned_artifact	Use the architecture-neutral training slot, a matched-budget classical angular-delay truncation, the full-CSI identity upper bound, or a returned artifact.

Differentiability (legacy trainable_params): framework=torch, gradient=full, trainable_params=True, exportable=True

The recipe declares the CSI-to-feedback interface; external training supplies its architecture.

Training capabilities: built_in_fine_tuning=False, portable_replacement=True

Portable trained-artifact ABI:

{
  "binding_params": {
    "artifact_entrypoint": "encoder",
    "artifact_manifest_path": "trained_artifact.yaml",
    "runtime": "learned_artifact"
  },
  "component_id": "encoder",
  "component_role": "csi_feedback_encoder",
  "entrypoint_id": "encoder",
  "inputs": {
    "csi_ri": {
      "dtype": "float32",
      "shape": [
        "batch",
        2,
        "tx_antenna",
        "subcarrier"
      ]
    }
  },
  "outputs": {
    "feedback_code": {
      "dtype": "float32",
      "shape": [
        "batch",
        "feedback_dimension"
      ]
    }
  },
  "required_operation_inputs": [
    "csi"
  ]
}

Backends:

Runner	Backends
benchmark_run	numpy, onnxruntime
dataset_capture	numpy, onnxruntime
differentiable_export	torch

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "npz",
  "checkpoint": "onnx",
  "tensor": "float32"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "identity_full_csi_encoder",
    "parameter_bindings": {
      "runtime": "identity"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "truncated_angular_delay_encoder",
    "parameter_bindings": {
      "runtime": "truncated_angular_delay"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "onnxruntime",
    "implementation": "portable_trained_artifact_encoder",
    "parameter_bindings": {
      "runtime": "learned_artifact"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "identity_full_csi_encoder",
    "parameter_bindings": {
      "runtime": "identity"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "truncated_angular_delay_encoder",
    "parameter_bindings": {
      "runtime": "truncated_angular_delay"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "onnxruntime",
    "implementation": "portable_trained_artifact_encoder",
    "parameter_bindings": {
      "runtime": "learned_artifact"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "trainable_csi_feedback_encoder_slot",
    "parameter_bindings": {
      "runtime": "training_interface"
    },
    "runner": "differentiable_export",
    "status": "implemented"
  }
]

model.csi_mrt_precoder

Name: MRT precoder from reconstructed CSI

Status: implemented

Inputs:

Name	Kind
reconstruction	channel.miso_ofdm_csi_reconstruction.numpy

Optional inputs:

None.

Outputs:

Name	Kind
precoder	channel.miso_ofdm_precoder.numpy

Parameters:

None.

Differentiability (legacy trainable_params): framework=torch, gradient=full, trainable_params=False, exportable=True

Per-subcarrier normalized MRT is differentiable with respect to reconstructed CSI.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy
dataset_capture	numpy
differentiable_export	torch

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "npz",
  "tensor": "float32 RI"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "default",
    "runner": "differentiable_export",
    "status": "implemented"
  }
]

model.deepjscc_external_decode

Name: DeepJSCC decoder replacement / returned-artifact slot

Status: implemented

Inputs:

Name	Kind
symbols	channel.rx_symbols.complex_numpy, channel.symbols.complex_numpy

Optional inputs:

None.

Outputs:

Name	Kind
images	image.batch.numpy

Parameters:

Name	Type	Required	Default / values	Description
artifact_entrypoint	string	no	default ``	Entrypoint in the registered artifact manifest; defaults to encoder or decoder by slot role.
artifact_manifest_path	string	no	default ``	Registered trained-artifact manifest implementing this DeepJSCC slot.
artifact_package_sha256	string	no	default ``	Required package identity of the frozen trained artifact.
call_style	string	no	default array_params values array_params, array, dict
callable	string	no	default ``
checkpoint_format	string	no	default noema_deepjscc_reference_cnn_npz_v1 values noema_deepjscc_reference_cnn_npz_v1
checkpoint_max_bytes	integer	no	default 67108864
checkpoint_path	string	no	default ``	Managed safe-NPZ DeepJSCC checkpoint used by runtime=learned_checkpoint.
checkpoint_sha256	string	no	default ``	Required lowercase SHA-256 of the frozen DeepJSCC checkpoint.
checkpoint_strict	boolean	no	default True
module	string	no	default ``
path	string	no	default ``
runtime	string	no	default training_interface values training_interface, external_callable, learned_checkpoint, learned_artifact	Declare an export-only interface, invoke a trusted callable, use the reference checkpoint runtime, or run a registered portable trained artifact.
symbol_channels	integer	no	default 16	Complex channel count declared by the managed checkpoint.

Differentiability (legacy trainable_params): framework=torch, gradient=full, trainable_params=True, exportable=True

This operation defines a portable decoder replacement boundary. When selected, its current implementation is omitted and the researcher-supplied module owns parameters and autograd behavior.

Training capabilities: built_in_fine_tuning=False, portable_replacement=True

Portable trained-artifact ABI:

{
  "binding_params": {
    "artifact_entrypoint": "decoder",
    "artifact_manifest_path": "trained_artifact.yaml",
    "runtime": "learned_artifact"
  },
  "component_id": "decoder",
  "component_role": "decoder",
  "entrypoint_id": "decoder",
  "inputs": {
    "symbols_ri": {
      "dtype": "float32",
      "shape": [
        "batch",
        "real_imag_channel",
        "symbol_height",
        "symbol_width"
      ]
    }
  },
  "outputs": {
    "reconstruction": {
      "dtype": "float32",
      "shape": [
        "batch",
        3,
        "height",
        "width"
      ]
    }
  },
  "required_operation_inputs": [
    "symbols"
  ]
}

Backends:

Runner	Backends
benchmark_run	external, torch, onnxruntime
dataset_capture	external, torch, onnxruntime
differentiable_export	torch

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "external",
    "implementation": "external_callable_runtime",
    "parameter_bindings": {
      "runtime": "external_callable"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "safe_npz_reference_cnn_runtime",
    "parameter_bindings": {
      "runtime": "learned_checkpoint"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "onnxruntime",
    "implementation": "portable_trained_artifact_runtime",
    "parameter_bindings": {
      "runtime": "learned_artifact"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "external",
    "implementation": "external_callable_runtime",
    "parameter_bindings": {
      "runtime": "external_callable"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "safe_npz_reference_cnn_runtime",
    "parameter_bindings": {
      "runtime": "learned_checkpoint"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "onnxruntime",
    "implementation": "portable_trained_artifact_runtime",
    "parameter_bindings": {
      "runtime": "learned_artifact"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "external_model_decoder_slot",
    "notes": "The recipe supplies the typed decoder boundary; the researcher supplies the model architecture and training method.",
    "parameter_bindings": {
      "runtime": "training_interface"
    },
    "runner": "differentiable_export",
    "status": "implemented"
  }
]

model.deepjscc_external_encode

Name: DeepJSCC encoder replacement / returned-artifact slot

Status: implemented

Inputs:

Name	Kind
images	image.batch.numpy

Optional inputs:

None.

Outputs:

Name	Kind
symbols	channel.symbols.complex_numpy

Parameters:

Name	Type	Required	Default / values	Description
artifact_entrypoint	string	no	default ``	Entrypoint in the registered artifact manifest; defaults to encoder or decoder by slot role.
artifact_manifest_path	string	no	default ``	Registered trained-artifact manifest implementing this DeepJSCC slot.
artifact_package_sha256	string	no	default ``	Required package identity of the frozen trained artifact.
call_style	string	no	default array_params values array_params, array, dict
callable	string	no	default ``
checkpoint_format	string	no	default noema_deepjscc_reference_cnn_npz_v1 values noema_deepjscc_reference_cnn_npz_v1
checkpoint_max_bytes	integer	no	default 67108864
checkpoint_path	string	no	default ``	Managed safe-NPZ DeepJSCC checkpoint used by runtime=learned_checkpoint.
checkpoint_sha256	string	no	default ``	Required lowercase SHA-256 of the frozen DeepJSCC checkpoint.
checkpoint_strict	boolean	no	default True
module	string	no	default ``
path	string	no	default ``
runtime	string	no	default training_interface values training_interface, external_callable, learned_checkpoint, learned_artifact	Declare an export-only interface, invoke a trusted callable, use the reference checkpoint runtime, or run a registered portable trained artifact.
symbol_channels	integer	no	default 16	Complex channel count declared by the managed checkpoint.

Differentiability (legacy trainable_params): framework=torch, gradient=full, trainable_params=True, exportable=True

This operation defines a portable encoder replacement boundary. When selected, its current implementation is omitted and the researcher-supplied module owns parameters and autograd behavior.

Training capabilities: built_in_fine_tuning=False, portable_replacement=True

Portable trained-artifact ABI:

{
  "binding_params": {
    "artifact_entrypoint": "encoder",
    "artifact_manifest_path": "trained_artifact.yaml",
    "runtime": "learned_artifact"
  },
  "component_id": "encoder",
  "component_role": "encoder",
  "entrypoint_id": "encoder",
  "inputs": {
    "images": {
      "dtype": "float32",
      "shape": [
        "batch",
        3,
        "height",
        "width"
      ]
    }
  },
  "outputs": {
    "symbols_ri": {
      "dtype": "float32",
      "shape": [
        "batch",
        "real_imag_channel",
        "symbol_height",
        "symbol_width"
      ]
    }
  },
  "required_operation_inputs": [
    "images"
  ]
}

Backends:

Runner	Backends
benchmark_run	external, torch, onnxruntime
dataset_capture	external, torch, onnxruntime
differentiable_export	torch

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "external",
    "implementation": "external_callable_runtime",
    "parameter_bindings": {
      "runtime": "external_callable"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "safe_npz_reference_cnn_runtime",
    "parameter_bindings": {
      "runtime": "learned_checkpoint"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "onnxruntime",
    "implementation": "portable_trained_artifact_runtime",
    "parameter_bindings": {
      "runtime": "learned_artifact"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "external",
    "implementation": "external_callable_runtime",
    "parameter_bindings": {
      "runtime": "external_callable"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "safe_npz_reference_cnn_runtime",
    "parameter_bindings": {
      "runtime": "learned_checkpoint"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "onnxruntime",
    "implementation": "portable_trained_artifact_runtime",
    "parameter_bindings": {
      "runtime": "learned_artifact"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "external_model_encoder_slot",
    "notes": "The recipe supplies the typed encoder boundary; the researcher supplies the model architecture and training method.",
    "parameter_bindings": {
      "runtime": "training_interface"
    },
    "runner": "differentiable_export",
    "status": "implemented"
  }
]

model.diffusers_autoencoderkl_decode

Name: Diffusers AutoencoderKL continuous latents decoder to image batch

Status: implemented

Inputs:

Name	Kind
latents	semantic.latents.numpy

Optional inputs:

None.

Outputs:

Name	Kind
images	image.batch.numpy

Parameters:

Name	Type	Required	Default / values	Description
device	string	no	default cpu
model_id	string	no	default stabilityai/sd-vae-ft-mse	Matched to the encoder model id by the dashboard.
revision	string	no	default ``	Full 40-character model-repository commit; required for remote models and matched to encoder evidence.
scale_latents	boolean	no	default True	Inverse of the encoder latent scaling.
scaling_factor	number	no	default 0.18215	Matched to the encoder scaling factor by the dashboard.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

model.diffusers_autoencoderkl_encode

Name: Diffusers AutoencoderKL image encoder to continuous latents

Status: implemented

Inputs:

Name	Kind
images	image.batch.numpy

Optional inputs:

None.

Outputs:

Name	Kind
latents	semantic.latents.numpy

Parameters:

Name	Type	Required	Default / values	Description
device	string	no	default cpu
model_id	string	no	default stabilityai/sd-vae-ft-mse	Diffusers AutoencoderKL repository id used by both encoder and decoder.
revision	string	no	default ``	Full 40-character model-repository commit; required for remote models and matched by the decoder.
sample_mode	string	no	default mean values mean, sample	mean is deterministic; sample draws from the VAE latent distribution and can vary between runs.
scale_latents	boolean	no	default True	Applies the Stable Diffusion latent scale before transmission; decoder applies the inverse.
scaling_factor	number	no	default 0.18215	Latent scale factor shared by encoder and decoder. Keep this matched to the model config.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

model.diffusers_vqmodel_decode

Name: Diffusers VQModel quantized semantic latents decoder to image batch

Status: implemented

Inputs:

Name	Kind
latents	semantic.latents.numpy

Optional inputs:

None.

Outputs:

Name	Kind
images	image.batch.numpy

Parameters:

Name	Type	Required	Default / values	Description
device	string	no	default cpu
model_id	string	no	default CompVis/ldm-celebahq-256
revision	string	no	default ``	Full 40-character model-repository commit; required for remote models and matched to encoder evidence.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

model.diffusers_vqmodel_encode

Name: Diffusers VQModel image encoder to quantized semantic latents

Status: implemented

Inputs:

Name	Kind
images	image.batch.numpy

Optional inputs:

None.

Outputs:

Name	Kind
latents	semantic.latents.numpy

Parameters:

Name	Type	Required	Default / values	Description
device	string	no	default cpu
model_id	string	no	default CompVis/ldm-celebahq-256
revision	string	no	default ``	Full 40-character model-repository commit; required for remote models and matched by the decoder.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

model.eflic_aoti_decode

Name: EF-LIC AOT Inductor fixed-length payload bits decoder

Status: implemented

Inputs:

Name	Kind
bits	channel.payload_bits.numpy, channel.bits.numpy
model	model.aot_inductor.bundle

Optional inputs:

None.

Outputs:

Name	Kind
images	image.batch.numpy

Parameters:

Name	Type	Required	Default / values	Description
device	string	no	default cpu
on_error	string	no	default fail values fail, zeros

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

Entropy/payload info:

{
  "algorithm": "fixed-length VQ/RVQ index packing",
  "coder": "none",
  "implementation": "NumPy bit unpacking + AOT Inductor EF-LIC decompress wrapper",
  "language": "Python/NumPy + PyTorch AOT Inductor",
  "note": "EF-LIC removes entropy coding; transmitted bits are fixed-length packed VQ/RVQ indices selected by force_ind."
}

model.eflic_aoti_decode_indices

Name: EF-LIC AOT Inductor decoder from VQ/RVQ indices

Status: implemented

Inputs:

Name	Kind
indices	semantic.indices.numpy
model	model.aot_inductor.bundle

Optional inputs:

None.

Outputs:

Name	Kind
images	image.batch.numpy

Parameters:

Name	Type	Required	Default / values	Description
device	string	no	default cpu
on_error	string	no	default fail values fail, zeros

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

Entropy/payload info:

{
  "algorithm": "fixed-length VQ/RVQ index packing",
  "coder": "none",
  "implementation": "AOT Inductor EF-LIC decompress wrapper",
  "language": "PyTorch AOT Inductor",
  "note": "EF-LIC removes entropy coding; transmitted bits are fixed-length packed VQ/RVQ indices selected by force_ind."
}

model.eflic_aoti_encode

Name: EF-LIC AOT Inductor encoder to fixed-length payload bits

Status: implemented

Inputs:

Name	Kind
images	image.batch.numpy
model	model.aot_inductor.bundle

Optional inputs:

None.

Outputs:

Name	Kind
bits	channel.payload_bits.numpy

Parameters:

Name	Type	Required	Default / values	Description
device	string	no	default cpu

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

Entropy/payload info:

{
  "algorithm": "fixed-length VQ/RVQ index packing",
  "coder": "none",
  "implementation": "AOT Inductor EF-LIC compress wrapper + NumPy bit packing",
  "language": "PyTorch AOT Inductor + Python/NumPy",
  "note": "EF-LIC removes entropy coding; transmitted bits are fixed-length packed VQ/RVQ indices selected by force_ind."
}

model.eflic_aoti_encode_indices

Name: EF-LIC AOT Inductor encoder to VQ/RVQ indices

Status: implemented

Inputs:

Name	Kind
images	image.batch.numpy
model	model.aot_inductor.bundle

Optional inputs:

None.

Outputs:

Name	Kind
indices	semantic.indices.numpy

Parameters:

Name	Type	Required	Default / values	Description
device	string	no	default cpu

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

Entropy/payload info:

{
  "algorithm": "fixed-length VQ/RVQ index packing",
  "coder": "none",
  "implementation": "AOT Inductor EF-LIC compress wrapper",
  "language": "PyTorch AOT Inductor",
  "note": "EF-LIC removes entropy coding; transmitted bits are fixed-length packed VQ/RVQ indices selected by force_ind."
}

model.eflic_bits_to_indices

Name: EF-LIC payload decoder from bits to VQ/RVQ indices

Status: implemented

Inputs:

Name	Kind
bits	channel.payload_bits.numpy, channel.bits.numpy

Optional inputs:

None.

Outputs:

Name	Kind
indices	semantic.indices.numpy

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

Entropy/payload info:

{
  "algorithm": "fixed-length VQ/RVQ index packing",
  "coder": "none",
  "implementation": "EF-LIC official inference model + NumPy bit packing",
  "language": "Python/PyTorch + NumPy",
  "note": "EF-LIC removes entropy coding; transmitted bits are fixed-length packed VQ/RVQ indices selected by force_ind."
}

model.eflic_decode

Name: EF-LIC fixed-length payload bits decoder to image batch

Status: implemented

Inputs:

Name	Kind
bits	channel.payload_bits.numpy, channel.bits.numpy

Optional inputs:

None.

Outputs:

Name	Kind
images	image.batch.numpy

Parameters:

Name	Type	Required	Default / values	Description
auto_setup	boolean	no	default False	Opt in to downloading EF_LIC.py. An expected_model_sha256 is mandatory when enabled.
checkpoint	string	no	default .noema/checkpoints/eflic/checkpoint.pth.tar	Local path to the pretrained EF-LIC checkpoint.pth.tar file.
checkpoint_url	string	no	default https://drive.google.com/file/d/1XrfmdUx0nFFBg9_ToVzz-A2jFZ5qiGR6/view?usp=sharing	Official checkpoint page. Browser download is often more reliable than automated Google Drive download.
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
device	string	no	default cpu
expected_checkpoint_sha256	string	no		Optional expected SHA-256 for the local EF-LIC checkpoint; release recipes must supply it.
expected_model_sha256	string	no		Required SHA-256 for EF_LIC.py when auto setup is enabled; also verifies an existing local file when supplied.
force_ind	integer	no	default 2	EF-LIC rate point. The official inference script evaluates force_ind values 0, 1, 2, 3, and 4.
model_url	string	no	default https://raw.githubusercontent.com/SevenCTHU/EF-LIC/main/EF_LIC.py	Raw EF_LIC.py URL used when auto setup is enabled and repo_path is missing the file.
on_error	string	no	default fail values fail, zeros
pad_to_multiple	integer	no	default 64	EF-LIC pads inputs to multiples of 64 using replicate padding.
repo_path	string	no	default .noema/upstreams/EF-LIC	Local path containing the official EF_LIC.py inference file.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

Entropy/payload info:

{
  "algorithm": "fixed-length VQ/RVQ index packing",
  "coder": "none",
  "implementation": "EF-LIC official inference model + NumPy bit packing",
  "language": "Python/PyTorch + NumPy",
  "note": "EF-LIC removes entropy coding; transmitted bits are fixed-length packed VQ/RVQ indices selected by force_ind."
}

model.eflic_decode_indices

Name: EF-LIC decoder from VQ/RVQ indices to image batch

Status: implemented

Inputs:

Name	Kind
indices	semantic.indices.numpy

Optional inputs:

None.

Outputs:

Name	Kind
images	image.batch.numpy

Parameters:

Name	Type	Required	Default / values	Description
auto_setup	boolean	no	default False	Opt in to downloading EF_LIC.py. An expected_model_sha256 is mandatory when enabled.
checkpoint	string	no	default .noema/checkpoints/eflic/checkpoint.pth.tar	Local path to the pretrained EF-LIC checkpoint.pth.tar file.
checkpoint_url	string	no	default https://drive.google.com/file/d/1XrfmdUx0nFFBg9_ToVzz-A2jFZ5qiGR6/view?usp=sharing	Official checkpoint page. Browser download is often more reliable than automated Google Drive download.
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
device	string	no	default cpu
expected_checkpoint_sha256	string	no		Optional expected SHA-256 for the local EF-LIC checkpoint; release recipes must supply it.
expected_model_sha256	string	no		Required SHA-256 for EF_LIC.py when auto setup is enabled; also verifies an existing local file when supplied.
force_ind	integer	no	default 2	EF-LIC rate point. The official inference script evaluates force_ind values 0, 1, 2, 3, and 4.
model_url	string	no	default https://raw.githubusercontent.com/SevenCTHU/EF-LIC/main/EF_LIC.py	Raw EF_LIC.py URL used when auto setup is enabled and repo_path is missing the file.
on_error	string	no	default fail values fail, zeros
pad_to_multiple	integer	no	default 64	EF-LIC pads inputs to multiples of 64 using replicate padding.
repo_path	string	no	default .noema/upstreams/EF-LIC	Local path containing the official EF_LIC.py inference file.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

Entropy/payload info:

{
  "algorithm": "fixed-length VQ/RVQ index packing",
  "coder": "none",
  "implementation": "EF-LIC official inference model + NumPy bit packing",
  "language": "Python/PyTorch + NumPy",
  "note": "EF-LIC removes entropy coding; transmitted bits are fixed-length packed VQ/RVQ indices selected by force_ind."
}

model.eflic_encode

Name: EF-LIC encoder to fixed-length payload bits

Status: implemented

Inputs:

Name	Kind
images	image.batch.numpy

Optional inputs:

None.

Outputs:

Name	Kind
bits	channel.payload_bits.numpy

Parameters:

Name	Type	Required	Default / values	Description
auto_setup	boolean	no	default False	Opt in to downloading EF_LIC.py. An expected_model_sha256 is mandatory when enabled.
checkpoint	string	no	default .noema/checkpoints/eflic/checkpoint.pth.tar	Local path to the pretrained EF-LIC checkpoint.pth.tar file.
checkpoint_url	string	no	default https://drive.google.com/file/d/1XrfmdUx0nFFBg9_ToVzz-A2jFZ5qiGR6/view?usp=sharing	Official checkpoint page. Browser download is often more reliable than automated Google Drive download.
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
device	string	no	default cpu
expected_checkpoint_sha256	string	no		Optional expected SHA-256 for the local EF-LIC checkpoint; release recipes must supply it.
expected_model_sha256	string	no		Required SHA-256 for EF_LIC.py when auto setup is enabled; also verifies an existing local file when supplied.
force_ind	integer	no	default 2	EF-LIC rate point. The official inference script evaluates force_ind values 0, 1, 2, 3, and 4.
model_url	string	no	default https://raw.githubusercontent.com/SevenCTHU/EF-LIC/main/EF_LIC.py	Raw EF_LIC.py URL used when auto setup is enabled and repo_path is missing the file.
pad_to_multiple	integer	no	default 64	EF-LIC pads inputs to multiples of 64 using replicate padding.
repo_path	string	no	default .noema/upstreams/EF-LIC	Local path containing the official EF_LIC.py inference file.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

Entropy/payload info:

{
  "algorithm": "fixed-length VQ/RVQ index packing",
  "coder": "none",
  "implementation": "EF-LIC official inference model + NumPy bit packing",
  "language": "Python/PyTorch + NumPy",
  "note": "EF-LIC removes entropy coding; transmitted bits are fixed-length packed VQ/RVQ indices selected by force_ind."
}

model.eflic_encode_indices

Name: EF-LIC encoder to VQ/RVQ indices

Status: implemented

Inputs:

Name	Kind
images	image.batch.numpy

Optional inputs:

None.

Outputs:

Name	Kind
indices	semantic.indices.numpy

Parameters:

Name	Type	Required	Default / values	Description
auto_setup	boolean	no	default False	Opt in to downloading EF_LIC.py. An expected_model_sha256 is mandatory when enabled.
checkpoint	string	no	default .noema/checkpoints/eflic/checkpoint.pth.tar	Local path to the pretrained EF-LIC checkpoint.pth.tar file.
checkpoint_url	string	no	default https://drive.google.com/file/d/1XrfmdUx0nFFBg9_ToVzz-A2jFZ5qiGR6/view?usp=sharing	Official checkpoint page. Browser download is often more reliable than automated Google Drive download.
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
device	string	no	default cpu
expected_checkpoint_sha256	string	no		Optional expected SHA-256 for the local EF-LIC checkpoint; release recipes must supply it.
expected_model_sha256	string	no		Required SHA-256 for EF_LIC.py when auto setup is enabled; also verifies an existing local file when supplied.
force_ind	integer	no	default 2	EF-LIC rate point. The official inference script evaluates force_ind values 0, 1, 2, 3, and 4.
model_url	string	no	default https://raw.githubusercontent.com/SevenCTHU/EF-LIC/main/EF_LIC.py	Raw EF_LIC.py URL used when auto setup is enabled and repo_path is missing the file.
pad_to_multiple	integer	no	default 64	EF-LIC pads inputs to multiples of 64 using replicate padding.
repo_path	string	no	default .noema/upstreams/EF-LIC	Local path containing the official EF_LIC.py inference file.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

Entropy/payload info:

{
  "algorithm": "fixed-length VQ/RVQ index packing",
  "coder": "none",
  "implementation": "EF-LIC official inference model + NumPy bit packing",
  "language": "Python/PyTorch + NumPy",
  "note": "EF-LIC removes entropy coding; transmitted bits are fixed-length packed VQ/RVQ indices selected by force_ind."
}

model.eflic_export_aoti

Name: Compile EF-LIC fixed-shape wrappers with AOT Inductor

Status: implemented

Inputs:

Name	Kind
images	image.batch.numpy

Optional inputs:

None.

Outputs:

Name	Kind
model	model.aot_inductor.bundle

Parameters:

Name	Type	Required	Default / values	Description
auto_setup	boolean	no	default False	Opt in to downloading EF_LIC.py. An expected_model_sha256 is mandatory when enabled.
checkpoint	string	no	default .noema/checkpoints/eflic/checkpoint.pth.tar	Local path to the pretrained EF-LIC checkpoint.pth.tar file.
checkpoint_url	string	no	default https://drive.google.com/file/d/1XrfmdUx0nFFBg9_ToVzz-A2jFZ5qiGR6/view?usp=sharing	Official checkpoint page. Browser download is often more reliable than automated Google Drive download.
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
device	string	no	default cpu
expected_checkpoint_sha256	string	no		Optional expected SHA-256 for the local EF-LIC checkpoint; release recipes must supply it.
expected_model_sha256	string	no		Required SHA-256 for EF_LIC.py when auto setup is enabled; also verifies an existing local file when supplied.
force_ind	integer	no	default 2	EF-LIC rate point. The official inference script evaluates force_ind values 0, 1, 2, 3, and 4.
model_url	string	no	default https://raw.githubusercontent.com/SevenCTHU/EF-LIC/main/EF_LIC.py	Raw EF_LIC.py URL used when auto setup is enabled and repo_path is missing the file.
pad_to_multiple	integer	no	default 64	EF-LIC pads inputs to multiples of 64 using replicate padding.
repo_path	string	no	default .noema/upstreams/EF-LIC	Local path containing the official EF_LIC.py inference file.
validate_export	boolean	no	default True

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

model.eflic_export_onnx

Name: Export EF-LIC fixed-shape ONNX Runtime bundle

Status: implemented

Inputs:

Name	Kind
images	image.batch.numpy

Optional inputs:

None.

Outputs:

Name	Kind
model	model.onnx.bundle

Parameters:

Name	Type	Required	Default / values	Description
auto_setup	boolean	no	default False	Opt in to downloading EF_LIC.py. An expected_model_sha256 is mandatory when enabled.
checkpoint	string	no	default .noema/checkpoints/eflic/checkpoint.pth.tar	Local path to the pretrained EF-LIC checkpoint.pth.tar file.
checkpoint_url	string	no	default https://drive.google.com/file/d/1XrfmdUx0nFFBg9_ToVzz-A2jFZ5qiGR6/view?usp=sharing	Official checkpoint page. Browser download is often more reliable than automated Google Drive download.
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
device	string	no	default cpu
expected_checkpoint_sha256	string	no		Optional expected SHA-256 for the local EF-LIC checkpoint; release recipes must supply it.
expected_model_sha256	string	no		Required SHA-256 for EF_LIC.py when auto setup is enabled; also verifies an existing local file when supplied.
force_ind	integer	no	default 2	EF-LIC rate point. The official inference script evaluates force_ind values 0, 1, 2, 3, and 4.
model_url	string	no	default https://raw.githubusercontent.com/SevenCTHU/EF-LIC/main/EF_LIC.py	Raw EF_LIC.py URL used when auto setup is enabled and repo_path is missing the file.
opset	integer	no	default 18
pad_to_multiple	integer	no	default 64	EF-LIC pads inputs to multiples of 64 using replicate padding.
repo_path	string	no	default .noema/upstreams/EF-LIC	Local path containing the official EF_LIC.py inference file.
validate_export	boolean	no	default True

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

model.eflic_indices_to_bits

Name: EF-LIC payload encoder from VQ/RVQ indices to bits

Status: implemented

Inputs:

Name	Kind
indices	semantic.indices.numpy

Optional inputs:

None.

Outputs:

Name	Kind
bits	channel.payload_bits.numpy

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

Entropy/payload info:

{
  "algorithm": "fixed-length VQ/RVQ index packing",
  "coder": "none",
  "implementation": "EF-LIC official inference model + NumPy bit packing",
  "language": "Python/PyTorch + NumPy",
  "note": "EF-LIC removes entropy coding; transmitted bits are fixed-length packed VQ/RVQ indices selected by force_ind."
}

model.eflic_onnx_decode

Name: EF-LIC ONNX Runtime fixed-length payload bits decoder

Status: implemented

Inputs:

Name	Kind
bits	channel.payload_bits.numpy, channel.bits.numpy
model	model.onnx.bundle

Optional inputs:

None.

Outputs:

Name	Kind
images	image.batch.numpy

Parameters:

Name	Type	Required	Default / values	Description
on_error	string	no	default fail values fail, zeros
provider	string	no	default CPUExecutionProvider

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

Entropy/payload info:

{
  "algorithm": "fixed-length VQ/RVQ index packing",
  "coder": "none",
  "implementation": "NumPy bit unpacking + ONNX Runtime EF-LIC decompress wrapper",
  "language": "Python/NumPy + ONNX Runtime",
  "note": "EF-LIC removes entropy coding; transmitted bits are fixed-length packed VQ/RVQ indices selected by force_ind."
}

model.eflic_onnx_decode_indices

Name: EF-LIC ONNX Runtime decoder from VQ/RVQ indices

Status: implemented

Inputs:

Name	Kind
indices	semantic.indices.numpy
model	model.onnx.bundle

Optional inputs:

None.

Outputs:

Name	Kind
images	image.batch.numpy

Parameters:

Name	Type	Required	Default / values	Description
on_error	string	no	default fail values fail, zeros
provider	string	no	default CPUExecutionProvider

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

Entropy/payload info:

{
  "algorithm": "fixed-length VQ/RVQ index packing",
  "coder": "none",
  "implementation": "ONNX Runtime EF-LIC decompress wrapper",
  "language": "ONNX Runtime",
  "note": "EF-LIC removes entropy coding; transmitted bits are fixed-length packed VQ/RVQ indices selected by force_ind."
}

model.eflic_onnx_encode

Name: EF-LIC ONNX Runtime encoder to fixed-length payload bits

Status: implemented

Inputs:

Name	Kind
images	image.batch.numpy
model	model.onnx.bundle

Optional inputs:

None.

Outputs:

Name	Kind
bits	channel.payload_bits.numpy

Parameters:

Name	Type	Required	Default / values	Description
provider	string	no	default CPUExecutionProvider

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

Entropy/payload info:

{
  "algorithm": "fixed-length VQ/RVQ index packing",
  "coder": "none",
  "implementation": "ONNX Runtime EF-LIC compress wrapper + NumPy bit packing",
  "language": "ONNX Runtime + Python/NumPy",
  "note": "EF-LIC removes entropy coding; transmitted bits are fixed-length packed VQ/RVQ indices selected by force_ind."
}

model.eflic_onnx_encode_indices

Name: EF-LIC ONNX Runtime encoder to VQ/RVQ indices

Status: implemented

Inputs:

Name	Kind
images	image.batch.numpy
model	model.onnx.bundle

Optional inputs:

None.

Outputs:

Name	Kind
indices	semantic.indices.numpy

Parameters:

Name	Type	Required	Default / values	Description
provider	string	no	default CPUExecutionProvider

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

Entropy/payload info:

{
  "algorithm": "fixed-length VQ/RVQ index packing",
  "coder": "none",
  "implementation": "ONNX Runtime EF-LIC compress wrapper",
  "language": "ONNX Runtime",
  "note": "EF-LIC removes entropy coding; transmitted bits are fixed-length packed VQ/RVQ indices selected by force_ind."
}

model.eflic_openvino_decode

Name: EF-LIC OpenVINO fixed-length payload bits decoder

Status: implemented

Inputs:

Name	Kind
bits	channel.payload_bits.numpy, channel.bits.numpy
model	model.onnx.bundle

Optional inputs:

None.

Outputs:

Name	Kind
images	image.batch.numpy

Parameters:

Name	Type	Required	Default / values	Description
device	string	no	default CPU
on_error	string	no	default fail values fail, zeros

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

Entropy/payload info:

{
  "algorithm": "fixed-length VQ/RVQ index packing",
  "coder": "none",
  "implementation": "NumPy bit unpacking + OpenVINO EF-LIC decompress wrapper",
  "language": "Python/NumPy + OpenVINO",
  "note": "EF-LIC removes entropy coding; transmitted bits are fixed-length packed VQ/RVQ indices selected by force_ind."
}

model.eflic_openvino_decode_indices

Name: EF-LIC OpenVINO decoder from VQ/RVQ indices

Status: implemented

Inputs:

Name	Kind
indices	semantic.indices.numpy
model	model.onnx.bundle

Optional inputs:

None.

Outputs:

Name	Kind
images	image.batch.numpy

Parameters:

Name	Type	Required	Default / values	Description
device	string	no	default CPU
on_error	string	no	default fail values fail, zeros

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

Entropy/payload info:

{
  "algorithm": "fixed-length VQ/RVQ index packing",
  "coder": "none",
  "implementation": "OpenVINO EF-LIC decompress wrapper",
  "language": "OpenVINO",
  "note": "EF-LIC removes entropy coding; transmitted bits are fixed-length packed VQ/RVQ indices selected by force_ind."
}

model.eflic_openvino_encode

Name: EF-LIC OpenVINO encoder to fixed-length payload bits

Status: implemented

Inputs:

Name	Kind
images	image.batch.numpy
model	model.onnx.bundle

Optional inputs:

None.

Outputs:

Name	Kind
bits	channel.payload_bits.numpy

Parameters:

Name	Type	Required	Default / values	Description
device	string	no	default CPU

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

Entropy/payload info:

{
  "algorithm": "fixed-length VQ/RVQ index packing",
  "coder": "none",
  "implementation": "OpenVINO EF-LIC compress wrapper + NumPy bit packing",
  "language": "OpenVINO + Python/NumPy",
  "note": "EF-LIC removes entropy coding; transmitted bits are fixed-length packed VQ/RVQ indices selected by force_ind."
}

model.eflic_openvino_encode_indices

Name: EF-LIC OpenVINO encoder to VQ/RVQ indices

Status: implemented

Inputs:

Name	Kind
images	image.batch.numpy
model	model.onnx.bundle

Optional inputs:

None.

Outputs:

Name	Kind
indices	semantic.indices.numpy

Parameters:

Name	Type	Required	Default / values	Description
device	string	no	default CPU

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

Entropy/payload info:

{
  "algorithm": "fixed-length VQ/RVQ index packing",
  "coder": "none",
  "implementation": "OpenVINO EF-LIC compress wrapper",
  "language": "OpenVINO",
  "note": "EF-LIC removes entropy coding; transmitted bits are fixed-length packed VQ/RVQ indices selected by force_ind."
}

model.equal_power_allocator

Name: Equal-power allocation baseline

Status: implemented

Inputs:

Name	Kind
problem	ai_phy.resource_allocation_problem.numpy

Optional inputs:

None.

Outputs:

Name	Kind
decision	ai_phy.resource_allocation_decision.numpy

Parameters:

None.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy
dataset_capture	numpy
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

model.evc_decode

Name: EVC upstream payload bits decoder to image batch

Status: implemented

Inputs:

Name	Kind
bits	channel.payload_bits.numpy, channel.bits.numpy

Optional inputs:

None.

Outputs:

Name	Kind
images	image.batch.numpy

Parameters:

Name	Type	Required	Default / values	Description
auto_setup	boolean	no	default False	Opt in to fetching pinned upstream assets. repo_revision and expected_checkpoint_sha256 are mandatory when downloads are needed.
checkpoint	string	no	default .noema/checkpoints/evc/EVC_SS_MD.pth.tar	Local path to the pretrained checkpoint for this RD point.
checkpoint_preset	string	no	default evc_ss_md	Dashboard preset identifier for the selected upstream checkpoint.
checkpoint_url	string	no	default https://onedrive.live.com/download?cid=2866592D5C55DF8C&resid=2866592D5C55DF8C%211231&authkey=ANrIn85RgtBH2wM	Pretrained checkpoint URL used when auto setup is enabled.
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
device	string	no	default cpu
expected_checkpoint_sha256	string	no		Expected checkpoint SHA-256 required for automatic checkpoint download.
on_error	string	no	default fail values fail, zeros
repo_clone_path	string	no	default .noema/upstreams/DCVC	Local clone target. For EVC this is the DCVC root; repo_path points to DCVC-family/EVC inside it.
repo_path	string	no	default .noema/upstreams/DCVC/DCVC-family/EVC	Local path to the official upstream repository clone.
repo_revision	string	no		Immutable 40-character Git commit required for automatic repository setup.
repo_url	string	no	default https://github.com/microsoft/DCVC.git	Official upstream Git repository URL used when auto setup is enabled.
setup_timeout_s	integer	no	default 900	Maximum time for each upstream auto-setup clone/download operation.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

Entropy/payload info:

{
  "algorithm": "range Asymmetric Numeral System",
  "coder": "rANS",
  "implementation": "EVC MLCodec_rans pybind11 extension",
  "language": "C++ extension + Python/PyTorch",
  "note": "EVC's upstream EntropyCoder wraps MLCodec_rans.RansEncoder/RansDecoder."
}

model.evc_encode

Name: EVC upstream checkpoint encoder to payload bits

Status: implemented

Inputs:

Name	Kind
images	image.batch.numpy

Optional inputs:

None.

Outputs:

Name	Kind
bits	channel.payload_bits.numpy

Parameters:

Name	Type	Required	Default / values	Description
auto_setup	boolean	no	default False	Opt in to fetching pinned upstream assets. repo_revision and expected_checkpoint_sha256 are mandatory when downloads are needed.
checkpoint	string	no	default .noema/checkpoints/evc/EVC_SS_MD.pth.tar	Local path to the pretrained checkpoint for this RD point.
checkpoint_preset	string	no	default evc_ss_md	Dashboard preset identifier for the selected upstream checkpoint.
checkpoint_url	string	no	default https://onedrive.live.com/download?cid=2866592D5C55DF8C&resid=2866592D5C55DF8C%211231&authkey=ANrIn85RgtBH2wM	Pretrained checkpoint URL used when auto setup is enabled.
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
device	string	no	default cpu
ec_thread	boolean	no	default False
expected_checkpoint_sha256	string	no		Expected checkpoint SHA-256 required for automatic checkpoint download.
model_name	string	no	default EVC_SS values EVC_LL, EVC_ML, EVC_SL, EVC_LM, EVC_LS, EVC_MM, EVC_SS, Scale_EVC_SL, Scale_EVC_SS
rate_idx	integer	no	default 0
repo_clone_path	string	no	default .noema/upstreams/DCVC	Local clone target. For EVC this is the DCVC root; repo_path points to DCVC-family/EVC inside it.
repo_path	string	no	default .noema/upstreams/DCVC/DCVC-family/EVC	Local path to the official upstream repository clone.
repo_revision	string	no		Immutable 40-character Git commit required for automatic repository setup.
repo_url	string	no	default https://github.com/microsoft/DCVC.git	Official upstream Git repository URL used when auto setup is enabled.
setup_timeout_s	integer	no	default 900	Maximum time for each upstream auto-setup clone/download operation.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

Entropy/payload info:

{
  "algorithm": "range Asymmetric Numeral System",
  "coder": "rANS",
  "implementation": "EVC MLCodec_rans pybind11 extension",
  "language": "C++ extension + Python/PyTorch",
  "note": "EVC's upstream EntropyCoder wraps MLCodec_rans.RansEncoder/RansDecoder."
}

model.evc_export_onnx

Name: Convert EVC neural transforms to ONNX

Status: implemented

Inputs:

None.

Optional inputs:

None.

Outputs:

Name	Kind
model	model.onnx.bundle

Parameters:

Name	Type	Required	Default / values	Description
auto_setup	boolean	no	default False	Opt in to fetching pinned upstream assets. repo_revision and expected_checkpoint_sha256 are mandatory when downloads are needed.
checkpoint	string	no	default .noema/checkpoints/evc/EVC_SS_MD.pth.tar	Local path to the pretrained checkpoint for this RD point.
checkpoint_preset	string	no	default evc_ss_md	Dashboard preset identifier for the selected upstream checkpoint.
checkpoint_url	string	no	default https://onedrive.live.com/download?cid=2866592D5C55DF8C&resid=2866592D5C55DF8C%211231&authkey=ANrIn85RgtBH2wM	Pretrained checkpoint URL used when auto setup is enabled.
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
device	string	no	default cpu
ec_thread	boolean	no	default False
expected_checkpoint_sha256	string	no		Expected checkpoint SHA-256 required for automatic checkpoint download.
export_height	integer	no	default 64
export_width	integer	no	default 64
model_name	string	no	default EVC_SS values EVC_LL, EVC_ML, EVC_SL, EVC_LM, EVC_LS, EVC_MM, EVC_SS, Scale_EVC_SL, Scale_EVC_SS
opset	integer	no	default 18
rate_idx	integer	no	default 0
repo_clone_path	string	no	default .noema/upstreams/DCVC	Local clone target. For EVC this is the DCVC root; repo_path points to DCVC-family/EVC inside it.
repo_path	string	no	default .noema/upstreams/DCVC/DCVC-family/EVC	Local path to the official upstream repository clone.
repo_revision	string	no		Immutable 40-character Git commit required for automatic repository setup.
repo_url	string	no	default https://github.com/microsoft/DCVC.git	Official upstream Git repository URL used when auto setup is enabled.
setup_timeout_s	integer	no	default 900	Maximum time for each upstream auto-setup clone/download operation.
validate_export	boolean	no	default True

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

Entropy/payload info:

{
  "algorithm": "range Asymmetric Numeral System",
  "coder": "rANS",
  "implementation": "EVC MLCodec_rans pybind11 extension",
  "language": "C++ extension + Python/PyTorch",
  "note": "EVC's upstream EntropyCoder wraps MLCodec_rans.RansEncoder/RansDecoder."
}

model.evc_onnx_decode

Name: EVC ONNX Runtime payload bits decoder to image batch

Status: implemented

Inputs:

Name	Kind
bits	channel.payload_bits.numpy, channel.bits.numpy
model	model.onnx.bundle

Optional inputs:

None.

Outputs:

Name	Kind
images	image.batch.numpy

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
device	string	no	default cpu
on_error	string	no	default fail values fail, zeros
provider	string	no	default CPUExecutionProvider

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

Entropy/payload info:

{
  "algorithm": "range Asymmetric Numeral System",
  "coder": "rANS",
  "implementation": "ONNX Runtime EVC neural transforms + EVC MLCodec_rans entropy backend",
  "language": "ONNX Runtime + C++ extension + Python/PyTorch",
  "note": "EVC's upstream EntropyCoder wraps MLCodec_rans.RansEncoder/RansDecoder."
}

model.evc_onnx_encode

Name: EVC ONNX Runtime encoder to payload bits

Status: implemented

Inputs:

Name	Kind
images	image.batch.numpy
model	model.onnx.bundle

Optional inputs:

None.

Outputs:

Name	Kind
bits	channel.payload_bits.numpy

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
device	string	no	default cpu
provider	string	no	default CPUExecutionProvider

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

Entropy/payload info:

{
  "algorithm": "range Asymmetric Numeral System",
  "coder": "rANS",
  "implementation": "ONNX Runtime EVC neural transforms + EVC MLCodec_rans entropy backend",
  "language": "ONNX Runtime + C++ extension + Python/PyTorch",
  "note": "EVC's upstream EntropyCoder wraps MLCodec_rans.RansEncoder/RansDecoder."
}

model.external_decode_bits

Name: External payload bits decoder to image batch

Status: implemented

Inputs:

Name	Kind
bits	channel.payload_bits.numpy, channel.bits.numpy

Optional inputs:

None.

Outputs:

Name	Kind
images	image.batch.numpy

Parameters:

Name	Type	Required	Default / values	Description
bit_order	string	no	default big values big, little	Bit order used when converting packed bytes to and from the channel bit vector.
bit_storage	string	no	default unpacked_bits values unpacked_bits, packed_bytes	unpacked_bits means one array element is one channel bit. packed_bytes means the callable returns bytes that Noema expands with np.unpackbits before channel transmission.
call_style	string	no	default array_params values array_params, array, dict
callable	string	no	default ``
module	string	no	default ``
path	string	no	default ``

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

model.external_decode_indices

Name: External semantic indices decoder to image batch

Status: implemented

Inputs:

Name	Kind
indices	semantic.indices.numpy

Optional inputs:

None.

Outputs:

Name	Kind
images	image.batch.numpy

Parameters:

Name	Type	Required	Default / values	Description
call_style	string	no	default array_params values array_params, array, dict
callable	string	no	default ``
module	string	no	default ``
path	string	no	default ``

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

model.external_decode_latents

Name: External continuous semantic latents decoder to image batch

Status: implemented

Inputs:

Name	Kind
latents	semantic.latents.numpy

Optional inputs:

None.

Outputs:

Name	Kind
images	image.batch.numpy

Parameters:

Name	Type	Required	Default / values	Description
call_style	string	no	default array_params values array_params, array, dict
callable	string	no	default ``
module	string	no	default ``
path	string	no	default ``

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

model.external_encode_bits

Name: External image encoder to payload bits

Status: implemented

Inputs:

Name	Kind
images	image.batch.numpy

Optional inputs:

None.

Outputs:

Name	Kind
bits	channel.payload_bits.numpy

Parameters:

Name	Type	Required	Default / values	Description
bit_order	string	no	default big values big, little	Bit order used when converting packed bytes to and from the channel bit vector.
bit_storage	string	no	default unpacked_bits values unpacked_bits, packed_bytes	unpacked_bits means one array element is one channel bit. packed_bytes means the callable returns bytes that Noema expands with np.unpackbits before channel transmission.
call_style	string	no	default array_params values array_params, array, dict
callable	string	no	default ``
module	string	no	default ``
path	string	no	default ``

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

model.external_encode_indices

Name: External image encoder to semantic indices

Status: implemented

Inputs:

Name	Kind
images	image.batch.numpy

Optional inputs:

None.

Outputs:

Name	Kind
indices	semantic.indices.numpy

Parameters:

Name	Type	Required	Default / values	Description
call_style	string	no	default array_params values array_params, array, dict
callable	string	no	default ``
module	string	no	default ``
path	string	no	default ``

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

model.external_encode_latents

Name: External image encoder to continuous semantic latents

Status: implemented

Inputs:

Name	Kind
images	image.batch.numpy

Optional inputs:

None.

Outputs:

Name	Kind
latents	semantic.latents.numpy

Parameters:

Name	Type	Required	Default / values	Description
call_style	string	no	default array_params values array_params, array, dict
callable	string	no	default ``
module	string	no	default ``
path	string	no	default ``

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

model.hpcm_decode

Name: HPCM upstream payload bits decoder to image batch

Status: implemented

Inputs:

Name	Kind
bits	channel.payload_bits.numpy, channel.bits.numpy

Optional inputs:

None.

Outputs:

Name	Kind
images	image.batch.numpy

Parameters:

Name	Type	Required	Default / values	Description
auto_setup	boolean	no	default False	Opt in to fetching pinned upstream assets. repo_revision and expected_checkpoint_sha256 are mandatory when downloads are needed.
checkpoint	string	no	default .noema/checkpoints/hpcm/hpcm_base_lambda0.013_mse.pth	Local path to the pretrained checkpoint for this RD point.
checkpoint_preset	string	no	default hpcm_base_l0.013_mse	Dashboard preset identifier for the selected upstream checkpoint.
checkpoint_url	string	no	default https://drive.google.com/file/d/1Snq7vkWQdApzCe-gK_V-WuRyMHQRL443/view?usp=drive_link	Pretrained checkpoint URL used when auto setup is enabled.
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
device	string	no	default cpu
expected_checkpoint_sha256	string	no		Expected checkpoint SHA-256 required for automatic checkpoint download.
on_error	string	no	default fail values fail, zeros
repo_clone_path	string	no	default .noema/upstreams/LIC-HPCM	Local clone target. For EVC this is the DCVC root; repo_path points to DCVC-family/EVC inside it.
repo_path	string	no	default .noema/upstreams/LIC-HPCM	Local path to the official upstream repository clone.
repo_revision	string	no		Immutable 40-character Git commit required for automatic repository setup.
repo_url	string	no	default https://github.com/lyq133/LIC-HPCM.git	Official upstream Git repository URL used when auto setup is enabled.
setup_timeout_s	integer	no	default 900	Maximum time for each upstream auto-setup clone/download operation.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

Entropy/payload info:

{
  "algorithm": "unbounded range Asymmetric Numeral System",
  "coder": "unbounded rANS",
  "implementation": "LIC-HPCM unbounded_ans pybind11 extension",
  "language": "C++ extension + Python/PyTorch",
  "note": "HPCM uses its upstream unbounded_rans arithmetic coder for real bitstream writing."
}

model.hpcm_encode

Name: HPCM upstream checkpoint encoder to payload bits

Status: implemented

Inputs:

Name	Kind
images	image.batch.numpy

Optional inputs:

None.

Outputs:

Name	Kind
bits	channel.payload_bits.numpy

Parameters:

Name	Type	Required	Default / values	Description
auto_setup	boolean	no	default False	Opt in to fetching pinned upstream assets. repo_revision and expected_checkpoint_sha256 are mandatory when downloads are needed.
checkpoint	string	no	default .noema/checkpoints/hpcm/hpcm_base_lambda0.013_mse.pth	Local path to the pretrained checkpoint for this RD point.
checkpoint_preset	string	no	default hpcm_base_l0.013_mse	Dashboard preset identifier for the selected upstream checkpoint.
checkpoint_url	string	no	default https://drive.google.com/file/d/1Snq7vkWQdApzCe-gK_V-WuRyMHQRL443/view?usp=drive_link	Pretrained checkpoint URL used when auto setup is enabled.
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
device	string	no	default cpu
expected_checkpoint_sha256	string	no		Expected checkpoint SHA-256 required for automatic checkpoint download.
model_name	string	no	default HPCM_Base values HPCM_Base, HPCM_Large, HPCM_Base_PhiContext, HPCM_1B
pad_to_multiple	integer	no	default 256
repo_clone_path	string	no	default .noema/upstreams/LIC-HPCM	Local clone target. For EVC this is the DCVC root; repo_path points to DCVC-family/EVC inside it.
repo_path	string	no	default .noema/upstreams/LIC-HPCM	Local path to the official upstream repository clone.
repo_revision	string	no		Immutable 40-character Git commit required for automatic repository setup.
repo_url	string	no	default https://github.com/lyq133/LIC-HPCM.git	Official upstream Git repository URL used when auto setup is enabled.
scale_table_levels	integer	no	default 60
setup_timeout_s	integer	no	default 900	Maximum time for each upstream auto-setup clone/download operation.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

Entropy/payload info:

{
  "algorithm": "unbounded range Asymmetric Numeral System",
  "coder": "unbounded rANS",
  "implementation": "LIC-HPCM unbounded_ans pybind11 extension",
  "language": "C++ extension + Python/PyTorch",
  "note": "HPCM uses its upstream unbounded_rans arithmetic coder for real bitstream writing."
}

model.isac_ofdm_allocator_adapter

Name: ISAC OFDM allocation adapter

Status: implemented

Inputs:

Name	Kind
problem	isac.ofdm_allocation_problem.numpy

Optional inputs:

None.

Outputs:

Name	Kind
decision	isac.ofdm_power_allocation.numpy

Parameters:

Name	Type	Required	Default / values	Description
artifact_entrypoint	string	no	default isac_allocator
artifact_manifest_path	string	no	default ``
artifact_package_sha256	string	no	default ``
mode	string	no	default equal_power values equal_power, communications_water_filling, scalarized_reference, learned_artifact

Differentiability (legacy trainable_params): framework=torch, gradient=full, trainable_params=True, exportable=True

The allocation boundary supports a portable power-simplex policy artifact.

Training capabilities: built_in_fine_tuning=False, portable_replacement=True

Portable trained-artifact ABI:

{
  "binding_params": {
    "artifact_entrypoint": "isac_allocator",
    "artifact_manifest_path": "trained_artifact.yaml",
    "mode": "learned_artifact"
  },
  "component_id": "allocator",
  "component_role": "joint_isac_ofdm_power_allocator",
  "entrypoint_id": "isac_allocator",
  "inputs": {
    "features": {
      "dtype": "float32",
      "shape": [
        "batch",
        "subcarrier",
        4
      ]
    }
  },
  "outputs": {
    "power": {
      "dtype": "float32",
      "shape": [
        "batch",
        "subcarrier"
      ]
    }
  },
  "required_operation_inputs": [
    "problem"
  ]
}

Backends:

Runner	Backends
benchmark_run	numpy, onnxruntime
dataset_capture	numpy, onnxruntime
differentiable_export	torch

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "equal_power",
    "parameter_bindings": {
      "mode": "equal_power"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "communication_water_filling",
    "parameter_bindings": {
      "mode": "communications_water_filling"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "scalarized_projected_gradient_oracle",
    "parameter_bindings": {
      "mode": "scalarized_reference"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "onnxruntime",
    "implementation": "portable_trained_artifact_runtime",
    "parameter_bindings": {
      "mode": "learned_artifact"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "reference_allocators",
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "label_free_scalarized_utility",
    "runner": "differentiable_export",
    "status": "implemented"
  }
]

model.jpeg_decode

Name: JPEG payload bits decoder to image batch

Status: implemented

Inputs:

Name	Kind
bits	channel.payload_bits.numpy, channel.bits.numpy

Optional inputs:

None.

Outputs:

Name	Kind
images	image.batch.numpy

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
on_error	string	no	default fail values fail, zeros, gray_image, erasure, report_outage	Use non-fail policies only when intentionally studying protected-link outages or corrupted entropy-coded payloads.

Differentiability (legacy trainable_params): framework=numpy, gradient=stop, trainable_params=False, exportable=False

Classical JPEG parsing, dequantization, and entropy decoding are not exposed as a differentiable training block.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

Entropy/payload info:

{
  "algorithm": "JPEG Huffman coding",
  "coder": "Huffman",
  "implementation": "Pillow JPEG backend",
  "language": "C/Python",
  "note": "Baseline JPEG comparisons normally use non-progressive JPEG without optimized Huffman-table search."
}

model.jpeg_encode

Name: JPEG image encoder to payload bits

Status: implemented

Inputs:

Name	Kind
images	image.batch.numpy

Optional inputs:

None.

Outputs:

Name	Kind
bits	channel.payload_bits.numpy

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
optimize	boolean	no	default False	Advanced. Extra entropy-coding pass for smaller files at the same quantization; leave off for a plain baseline JPEG comparison unless you report optimized JPEG.
progressive	boolean	no	default False	Advanced. Store the JPEG in progressive scan order for streaming/preview; leave off for the usual baseline sequential JPEG comparison unless explicitly studying progressive JPEG.
quality	integer	no	default 75	JPEG quality factor. Higher values keep more detail and use more bits.
subsampling	string	no	default 420 values keep, 444, 422, 420	Chroma subsampling. 444 preserves chroma best; 420 is the common higher-compression setting.
wire_format	string	no	default safe_json_base64 values safe_json_base64, compact_binary_v1	Artifact-safe JSON/base64 remains the compatibility default; compact_binary_v1 is the bounded communication wire format for transmitted-resource studies.

Differentiability (legacy trainable_params): framework=numpy, gradient=stop, trainable_params=False, exportable=False

Classical JPEG quantization and entropy coding are non-differentiable in Noema benchmark execution.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

Entropy/payload info:

{
  "algorithm": "JPEG Huffman coding",
  "coder": "Huffman",
  "implementation": "Pillow JPEG backend",
  "language": "C/Python",
  "note": "Baseline JPEG comparisons normally use non-progressive JPEG without optimized Huffman-table search."
}

model.leo_ntn_tracking_adapter

Name: LEO-NTN Doppler and beam-handover adapter

Status: implemented

Inputs:

Name	Kind
problem	ntn.tracking_history.numpy

Optional inputs:

Name	Kind
truth	ntn.future_state_truth.numpy

Outputs:

Name	Kind
decision	ntn.future_state_decision.numpy

Parameters:

Name	Type	Required	Default / values	Description
artifact_entrypoint	string	no	default ntn_tracker
artifact_manifest_path	string	no	default ``
artifact_package_sha256	string	no	default ``
mode	string	no	default linear_extrapolation values hold_last, linear_extrapolation, oracle_future, learned_artifact

Differentiability (legacy trainable_params): framework=torch, gradient=full, trainable_params=True, exportable=True

A bounded history tensor feeds a portable joint Doppler/beam predictor.

Training capabilities: built_in_fine_tuning=False, portable_replacement=True

Portable trained-artifact ABI:

{
  "binding_params": {
    "artifact_entrypoint": "ntn_tracker",
    "artifact_manifest_path": "trained_artifact.yaml",
    "mode": "learned_artifact"
  },
  "component_id": "tracker",
  "component_role": "leo_ntn_doppler_beam_tracker",
  "entrypoint_id": "ntn_tracker",
  "inputs": {
    "track_features": {
      "dtype": "float32",
      "shape": [
        "batch",
        "history",
        3
      ]
    }
  },
  "outputs": {
    "decision": {
      "dtype": "float32",
      "shape": [
        "batch",
        "decision"
      ]
    }
  },
  "required_operation_inputs": [
    "problem"
  ]
}

Backends:

Runner	Backends
benchmark_run	numpy, onnxruntime
dataset_capture	numpy, onnxruntime
differentiable_export	torch

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "hold_last_observation",
    "parameter_bindings": {
      "mode": "hold_last"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "linear_extrapolation",
    "parameter_bindings": {
      "mode": "linear_extrapolation"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "simulation_future_state",
    "parameter_bindings": {
      "mode": "oracle_future"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "onnxruntime",
    "implementation": "portable_trained_artifact_runtime",
    "parameter_bindings": {
      "mode": "learned_artifact"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "reference_trackers",
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "supervised_future_state_prediction",
    "runner": "differentiable_export",
    "status": "implemented"
  }
]

model.localization_adapter

Name: Localization/sensing adapter

Status: implemented

Inputs:

Name	Kind
problem	ai_phy.localization_problem.numpy, ai_phy.localization_observation.numpy

Optional inputs:

None.

Outputs:

Name	Kind
estimate	ai_phy.localization_estimate.numpy

Parameters:

Name	Type	Required	Default / values	Description
artifact_entrypoint	string	no	default localization_estimator
artifact_manifest_path	string	no	default ``	Registered schema-v2 trained artifact implementing the range-localization ABI.
artifact_package_sha256	string	no	default ``
mode	string	no	default regularized_trilateration values trilateration, regularized_trilateration, learned_artifact	Select a fixed trilateration method or a returned portable localizer.

Differentiability (legacy trainable_params): framework=torch, gradient=surrogate, trainable_params=True, exportable=True

The adapter accepts a portable learned range localizer while linear and centroid-regularized trilateration remain fixed baselines.

Training capabilities: built_in_fine_tuning=False, portable_replacement=True

Portable trained-artifact ABI:

{
  "binding_params": {
    "artifact_entrypoint": "localization_estimator",
    "artifact_manifest_path": "trained_artifact.yaml",
    "mode": "learned_artifact"
  },
  "component_id": "localizer",
  "component_role": "two_dimensional_range_localizer",
  "entrypoint_id": "localization_estimator",
  "inputs": {
    "anchors": {
      "dtype": "float32",
      "shape": [
        "batch",
        "anchor",
        2
      ]
    },
    "ranges": {
      "dtype": "float32",
      "shape": [
        "batch",
        "anchor"
      ]
    }
  },
  "outputs": {
    "positions": {
      "dtype": "float32",
      "shape": [
        "batch",
        2
      ]
    }
  },
  "required_operation_inputs": [
    "problem"
  ]
}

Backends:

Runner	Backends
benchmark_run	numpy, onnxruntime
dataset_capture	numpy, onnxruntime
differentiable_export	torch, sionna

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "linear_trilateration",
    "parameter_bindings": {
      "mode": "trilateration"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "centroid_regularized_trilateration",
    "parameter_bindings": {
      "mode": "regularized_trilateration"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "onnxruntime",
    "implementation": "portable_trained_artifact_runtime",
    "parameter_bindings": {
      "mode": "learned_artifact"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "linear_trilateration",
    "parameter_bindings": {
      "mode": "trilateration"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "centroid_regularized_trilateration",
    "parameter_bindings": {
      "mode": "regularized_trilateration"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "onnxruntime",
    "implementation": "portable_trained_artifact_runtime",
    "parameter_bindings": {
      "mode": "learned_artifact"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "trainable_localization_endpoint",
    "runner": "differentiable_export",
    "status": "implemented"
  },
  {
    "backend": "sionna",
    "implementation": "sionna_rt_localization_endpoint",
    "runner": "differentiable_export",
    "status": "planned"
  }
]

model.ls_channel_estimator

Name: LS channel estimator baseline

Status: implemented

Inputs:

Name	Kind
problem	ai_phy.channel_estimation_problem.numpy, ai_phy.channel_estimation_observation.numpy

Optional inputs:

None.

Outputs:

Name	Kind
estimate	ai_phy.channel_estimate.numpy

Parameters:

None.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Least-squares benchmark baseline is an artifact operation.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy
dataset_capture	numpy
differentiable_export	None

Equivalence:

{
  "tolerance": {
    "atol": 1e-07,
    "rtol": 1e-06
  },
  "type": "numerical"
}

Formats:

{
  "artifact": "npz",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

model.modulation_classifier_adapter

Name: Modulation classifier

Status: implemented

Inputs:

Name	Kind
observation	ai_phy.modulation_iq_frames.numpy

Optional inputs:

None.

Outputs:

Name	Kind
prediction	ai_phy.modulation_predictions.numpy

Parameters:

Name	Type	Required	Default / values	Description
artifact_entrypoint	string	no	default modulation_classifier
artifact_manifest_path	string	no	default ``	Schema-v2 trained artifact implementing the modulation-classifier ABI.
artifact_package_sha256	string	no	default ``
mode	string	no	default classical_cumulant values classical_cumulant, classical_likelihood, classical_evm, learned_artifact

Differentiability (legacy trainable_params): framework=torch, gradient=surrogate, trainable_params=True, exportable=True

The classifier slot accepts a portable learned artifact while classical modes remain reproducible baselines.

Training capabilities: built_in_fine_tuning=False, portable_replacement=True

Portable trained-artifact ABI:

{
  "binding_params": {
    "artifact_entrypoint": "modulation_classifier",
    "artifact_manifest_path": "trained_artifact.yaml",
    "mode": "learned_artifact"
  },
  "component_id": "classifier",
  "component_role": "modulation_classifier",
  "entrypoint_id": "modulation_classifier",
  "inputs": {
    "iq_ri": {
      "dtype": "float32",
      "shape": [
        "batch",
        "sample",
        2
      ]
    }
  },
  "outputs": {
    "class_logits": {
      "dtype": "float32",
      "shape": [
        "batch",
        3
      ]
    }
  },
  "required_operation_inputs": [
    "observation"
  ]
}

Backends:

Runner	Backends
benchmark_run	numpy, onnxruntime
dataset_capture	numpy, onnxruntime
differentiable_export	torch

Equivalence:

{
  "reason": "Classifiers are compared through the fixed class vocabulary and task metrics.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "npz",
  "checkpoint": "onnx",
  "tensor": "torch.Tensor"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "blind_differential_cumulant",
    "parameter_bindings": {
      "mode": "classical_cumulant"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "awgn_mixture_likelihood",
    "parameter_bindings": {
      "mode": "classical_likelihood"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "minimum_constellation_evm",
    "parameter_bindings": {
      "mode": "classical_evm"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "onnxruntime",
    "implementation": "portable_trained_artifact_runtime",
    "parameter_bindings": {
      "mode": "learned_artifact"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "blind_differential_cumulant",
    "parameter_bindings": {
      "mode": "classical_cumulant"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "awgn_mixture_likelihood",
    "parameter_bindings": {
      "mode": "classical_likelihood"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "onnxruntime",
    "implementation": "portable_trained_artifact_runtime",
    "parameter_bindings": {
      "mode": "learned_artifact"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "modulation_classification_training_endpoint",
    "runner": "differentiable_export",
    "status": "implemented"
  }
]

model.mrt_beamformer

Name: MRT/codebook beamformer baseline

Status: implemented

Inputs:

Name	Kind
problem	ai_phy.beamforming_problem.numpy

Optional inputs:

None.

Outputs:

Name	Kind
decision	ai_phy.beamforming_decision.numpy

Parameters:

None.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy
dataset_capture	numpy
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

model.music_aoa_estimator

Name: MUSIC AoA estimator baseline

Status: implemented

Inputs:

Name	Kind
problem	ai_phy.aoa_problem.numpy, ai_phy.aoa_observation.numpy

Optional inputs:

None.

Outputs:

Name	Kind
estimate	ai_phy.aoa_estimate.numpy

Parameters:

Name	Type	Required	Default / values	Description
grid_size	integer	no	default 721
grid_step_deg	number	no

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Classical sample-covariance eigendecomposition and grid search benchmark.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy
dataset_capture	numpy
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "npz",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

model.near_field_estimator_adapter

Name: Near-field range-angle estimator adapter

Status: implemented

Inputs:

Name	Kind
problem	near_field.array_observation.numpy

Optional inputs:

Name	Kind
truth	near_field.range_angle_truth.numpy

Outputs:

Name	Kind
estimate	near_field.range_angle_estimate.numpy

Parameters:

Name	Type	Required	Default / values	Description
artifact_entrypoint	string	no	default near_field_estimator
artifact_manifest_path	string	no	default ``
artifact_package_sha256	string	no	default ``
mode	string	no	default polar_codebook values far_field_steering, polar_codebook, oracle_focus, learned_artifact

Differentiability (legacy trainable_params): framework=torch, gradient=surrogate, trainable_params=True, exportable=True

The coherent array observation is exposed through a portable range-angle estimator ABI.

Training capabilities: built_in_fine_tuning=False, portable_replacement=True

Portable trained-artifact ABI:

{
  "binding_params": {
    "artifact_entrypoint": "near_field_estimator",
    "artifact_manifest_path": "trained_artifact.yaml",
    "mode": "learned_artifact"
  },
  "component_id": "estimator",
  "component_role": "near_field_range_angle_estimator",
  "entrypoint_id": "near_field_estimator",
  "inputs": {
    "array_ri": {
      "dtype": "float32",
      "shape": [
        "batch",
        "antenna",
        2
      ]
    }
  },
  "outputs": {
    "range_angle": {
      "dtype": "float32",
      "shape": [
        "batch",
        2
      ]
    }
  },
  "required_operation_inputs": [
    "problem"
  ]
}

Backends:

Runner	Backends
benchmark_run	numpy, onnxruntime
dataset_capture	numpy, onnxruntime
differentiable_export	torch

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "far_field_angle_grid",
    "parameter_bindings": {
      "mode": "far_field_steering"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "polar_range_angle_codebook",
    "parameter_bindings": {
      "mode": "polar_codebook"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "simulation_truth",
    "parameter_bindings": {
      "mode": "oracle_focus"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "onnxruntime",
    "implementation": "portable_trained_artifact_runtime",
    "parameter_bindings": {
      "mode": "learned_artifact"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "reference_estimators",
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "supervised_range_angle_estimation",
    "runner": "differentiable_export",
    "status": "implemented"
  }
]

model.resource_allocation_adapter

Name: Resource allocation adapter

Status: implemented

Inputs:

Name	Kind
problem	ai_phy.resource_allocation_problem.numpy

Optional inputs:

None.

Outputs:

Name	Kind
decision	ai_phy.resource_allocation_decision.numpy

Parameters:

Name	Type	Required	Default / values	Description
temperature	number	no	default 1.0

Differentiability (legacy trainable_params): framework=torch, gradient=surrogate, trainable_params=True, exportable=True

Surrogate-gradient metadata describes this adapter only when retained as downstream support. Portable replacement is not available until a trained-artifact ABI and runtime binding are implemented.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy
dataset_capture	numpy
differentiable_export	torch

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "weighted_gain_softmax_reference",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "weighted_gain_softmax_reference",
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "trainable_resource_policy_endpoint",
    "runner": "differentiable_export",
    "status": "implemented"
  }
]

model.symbol_power_allocator

Name: Budgeted symbol power allocator

Status: implemented

Inputs:

Name	Kind
symbols	channel.symbols.complex_numpy, channel.rx_symbols.complex_numpy

Optional inputs:

Name	Kind
bits	channel.payload_bits.numpy, channel.coded_bits.numpy, channel.bits.numpy
channel_state	channel.ofdm_channel_state.numpy

Outputs:

Name	Kind
allocation	channel.power_allocation.numpy
symbols	channel.symbols.complex_numpy

Parameters:

Name	Type	Required	Default / values	Description
allocation_contrast	number	no	default 0.6
allocation_lower_power_ratio	number	no	default 0.1	Minimum per-subcarrier power relative to the mean budget for bounded causal-AR water filling.
allocation_upper_power_ratio	number	no	default 1.9	Maximum per-subcarrier power relative to the mean budget for bounded causal-AR water filling.
artifact_entrypoint	string	no	default power_policy	Entrypoint in the registered artifact manifest used for allocator inference.
artifact_manifest_path	string	no	default ``	Registered trained-artifact manifest implementing the architecture-neutral allocator slot ABI.
artifact_package_sha256	string	no	default ``	Registry-independent digest of the complete trained-artifact package bound into the execution plan.
bit_loading_bpsk_min_snr_db	number	no	default 6.0
bit_loading_max_bits_per_symbol	integer	no	default 6 values 1, 2, 4, 6
bit_loading_qam16_min_snr_db	number	no	default 17.0
bit_loading_qam64_min_snr_db	number	no	default 23.0
bit_loading_qpsk_min_snr_db	number	no	default 10.0
budget_mode	string	no	default fixed_average values fixed_average, variable_average
checkpoint_format	string	no	default noema_csi_power_deepset_npz_v1 values noema_csi_power_deepset_npz_v1
checkpoint_max_bytes	integer	no	default 67108864
checkpoint_path	string	no	default ``	Frozen safe-NPZ CSI allocator checkpoint used by policy=learned_checkpoint.
checkpoint_sha256	string	no	default ``	Required lowercase SHA-256 of the frozen allocator checkpoint.
checkpoint_strict	boolean	no	default True
csi_gain_shrinkage	number	no	default 0.6	For uncertainty-shrunk water filling, blend observed gains toward their per-state mean before allocating power.
csi_prediction_gain_confidence	number	no	default 0.4	Blend complex-AR predicted gains toward their per-state frequency mean before water filling.
csi_prediction_horizon_ofdm_symbols	integer	no	default 0	For complex-AR prediction, forecast this many OFDM symbols past the newest causal CSI snapshot. Zero uses the CSI artifact’s declared feedback delay.
eps	number	no	default 1e-12
granularity	string	no	default global values global, per_symbol, per_subcarrier, per_stream
label	string	no	default tx_power_allocate
max_power	number	no	default 2.0
midpoint_snr_db	number	no	default 12.0
min_power	number	no	default 0.25
model_batch_size	integer	no	default 1024	Maximum number of independent channel-state examples evaluated in one learned-model inference call. Chunking preserves the complete experiment batch and output order.
policy	string	no	default snr_sigmoid values fixed, snr_sigmoid, water_filling, observed_csi_water_filling, robust_csi_water_filling, causal_ar_water_filling, causal_ar_box_water_filling, learned_checkpoint, learned_artifact	Power-allocation method. Model selection is shown only for learned policies.
power_unit	string	no	default normalized values normalized, mW
slope_db	number	no	default 4.0
snr_db	number	no	default 12.0	Fallback reference SNR used only by the SNR-adaptive sigmoid policy when no explicit channel state is connected.
stream_count	integer	no	default 1
subcarrier_count	integer	no	default 64
target_power	number	no	default 1.0	Normalized average complex-symbol power budget per resource element; the per-state sum constraint is this value times the subcarrier count.
transport_mode	string	no	default fixed_modulation values fixed_modulation, allocation_aware_bit_loading	Optionally convert the coded-bit stream into a CSI/allocation-aware OFDM modulation map.

Differentiability (legacy trainable_params): framework=torch, gradient=full, trainable_params=True, exportable=True

Continuous-symbol power allocation is differentiable. Noema can benchmark a frozen, CSI-conditioned artifact behind an architecture-neutral runtime ABI while model training remains outside the benchmark executor.

Training capabilities: built_in_fine_tuning=False, portable_replacement=True

Portable trained-artifact ABI:

{
  "binding_params": {
    "artifact_entrypoint": "power_policy",
    "artifact_manifest_path": "trained_artifact.yaml",
    "budget_mode": "fixed_average",
    "granularity": "per_subcarrier",
    "policy": "learned_artifact"
  },
  "component_id": "policy",
  "component_role": "power_policy",
  "constraint_adapter": "noema_exact_simplex_projection_v1",
  "entrypoint_id": "power_policy",
  "inputs": {
    "average_power_budget": {
      "dtype": "float32",
      "shape": [
        "batch",
        1
      ]
    },
    "channel_gain": {
      "dtype": "float32",
      "shape": [
        "batch",
        "subcarrier"
      ]
    },
    "noise_variance": {
      "dtype": "float32",
      "shape": [
        "batch",
        1
      ]
    }
  },
  "outputs": {
    "allocation_scores": {
      "dtype": "float32",
      "shape": [
        "batch",
        "subcarrier"
      ]
    }
  },
  "required_operation_inputs": [
    "channel_state"
  ]
}

Backends:

Runner	Backends
benchmark_run	numpy, onnxruntime
dataset_capture	numpy, onnxruntime
differentiable_export	torch

Equivalence:

{
  "reason": "Reference NumPy and Torch policies should select the same group power targets and scaled symbols for fixed parameters.",
  "tolerance": {
    "atol": 1e-06,
    "rtol": 1e-05
  },
  "type": "numerical"
}

Formats:

{
  "artifact": "npz",
  "checkpoint": "noema_csi_power_deepset_npz_v1",
  "tensor": "torch.Tensor"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "fixed_power_allocator",
    "parameter_bindings": {
      "policy": "fixed"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "snr_sigmoid_power_allocator",
    "parameter_bindings": {
      "policy": "snr_sigmoid"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "closed_form_water_filling_allocator",
    "parameter_bindings": {
      "policy": "water_filling"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "water_filling_on_observed_csi",
    "parameter_bindings": {
      "policy": "observed_csi_water_filling"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "correlation_shrunk_water_filling",
    "parameter_bindings": {
      "policy": "robust_csi_water_filling"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "causal_complex_ar_prediction_plus_water_filling",
    "parameter_bindings": {
      "policy": "causal_ar_water_filling"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "causal_complex_ar_prediction_plus_box_constrained_water_filling",
    "parameter_bindings": {
      "policy": "causal_ar_box_water_filling"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "onnxruntime",
    "implementation": "portable_trained_artifact_policy_plus_noema_projection",
    "parameter_bindings": {
      "policy": "learned_artifact"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "fixed_power_allocator",
    "parameter_bindings": {
      "policy": "fixed"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "snr_sigmoid_power_allocator",
    "parameter_bindings": {
      "policy": "snr_sigmoid"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "closed_form_water_filling_allocator",
    "parameter_bindings": {
      "policy": "water_filling"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "water_filling_on_observed_csi",
    "parameter_bindings": {
      "policy": "observed_csi_water_filling"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "correlation_shrunk_water_filling",
    "parameter_bindings": {
      "policy": "robust_csi_water_filling"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "causal_complex_ar_prediction_plus_water_filling",
    "parameter_bindings": {
      "policy": "causal_ar_water_filling"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "causal_complex_ar_prediction_plus_box_constrained_water_filling",
    "parameter_bindings": {
      "policy": "causal_ar_box_water_filling"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "onnxruntime",
    "implementation": "portable_trained_artifact_policy_plus_noema_projection",
    "parameter_bindings": {
      "policy": "learned_artifact"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "trainable_budgeted_symbol_allocator",
    "runner": "differentiable_export",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "safe_npz_csi_deepset_checkpoint",
    "parameter_bindings": {
      "policy": "learned_checkpoint"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "safe_npz_csi_deepset_checkpoint",
    "parameter_bindings": {
      "policy": "learned_checkpoint"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

model.tcm_decode

Name: TCM upstream payload bits decoder to image batch

Status: implemented

Inputs:

Name	Kind
bits	channel.payload_bits.numpy, channel.bits.numpy

Optional inputs:

None.

Outputs:

Name	Kind
images	image.batch.numpy

Parameters:

Name	Type	Required	Default / values	Description
auto_setup	boolean	no	default False	Opt in to fetching pinned upstream assets. repo_revision and expected_checkpoint_sha256 are mandatory when downloads are needed.
checkpoint	string	no	default .noema/checkpoints/tcm/tcm_N128_lambda0.05_mse.pth	Local path to the pretrained checkpoint for this RD point.
checkpoint_preset	string	no	default tcm_n128_l0.05_mse	Dashboard preset identifier for the selected upstream checkpoint.
checkpoint_url	string	no	default https://drive.google.com/file/d/1TK-CPiD2QwtWJqZoT_OyCtnxdQ7UNP56/view?usp=share_link	Pretrained checkpoint URL used when auto setup is enabled.
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
device	string	no	default cpu
expected_checkpoint_sha256	string	no		Expected checkpoint SHA-256 required for automatic checkpoint download.
on_error	string	no	default fail values fail, zeros
repo_clone_path	string	no	default .noema/upstreams/LIC_TCM	Local clone target. For EVC this is the DCVC root; repo_path points to DCVC-family/EVC inside it.
repo_path	string	no	default .noema/upstreams/LIC_TCM	Local path to the official upstream repository clone.
repo_revision	string	no		Immutable 40-character Git commit required for automatic repository setup.
repo_url	string	no	default https://github.com/jmliu206/LIC_TCM.git	Official upstream Git repository URL used when auto setup is enabled.
setup_timeout_s	integer	no	default 900	Maximum time for each upstream auto-setup clone/download operation.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

Entropy/payload info:

{
  "algorithm": "rANS",
  "coder": "CompressAI BufferedRansEncoder/RansDecoder",
  "implementation": "CompressAI ANS backend used by LIC_TCM",
  "language": "C++ extension + Python/PyTorch",
  "note": "TCM's upstream model imports compressai.ans.BufferedRansEncoder and RansDecoder for bitstream coding."
}

model.tcm_encode

Name: TCM upstream checkpoint encoder to payload bits

Status: implemented

Inputs:

Name	Kind
images	image.batch.numpy

Optional inputs:

None.

Outputs:

Name	Kind
bits	channel.payload_bits.numpy

Parameters:

Name	Type	Required	Default / values	Description
M	integer	no	default 320
N	integer	no	default 128
auto_setup	boolean	no	default False	Opt in to fetching pinned upstream assets. repo_revision and expected_checkpoint_sha256 are mandatory when downloads are needed.
checkpoint	string	no	default .noema/checkpoints/tcm/tcm_N128_lambda0.05_mse.pth	Local path to the pretrained checkpoint for this RD point.
checkpoint_preset	string	no	default tcm_n128_l0.05_mse	Dashboard preset identifier for the selected upstream checkpoint.
checkpoint_url	string	no	default https://drive.google.com/file/d/1TK-CPiD2QwtWJqZoT_OyCtnxdQ7UNP56/view?usp=share_link	Pretrained checkpoint URL used when auto setup is enabled.
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
device	string	no	default cpu
expected_checkpoint_sha256	string	no		Expected checkpoint SHA-256 required for automatic checkpoint download.
pad_to_multiple	integer	no	default 128
repo_clone_path	string	no	default .noema/upstreams/LIC_TCM	Local clone target. For EVC this is the DCVC root; repo_path points to DCVC-family/EVC inside it.
repo_path	string	no	default .noema/upstreams/LIC_TCM	Local path to the official upstream repository clone.
repo_revision	string	no		Immutable 40-character Git commit required for automatic repository setup.
repo_url	string	no	default https://github.com/jmliu206/LIC_TCM.git	Official upstream Git repository URL used when auto setup is enabled.
setup_timeout_s	integer	no	default 900	Maximum time for each upstream auto-setup clone/download operation.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

Entropy/payload info:

{
  "algorithm": "rANS",
  "coder": "CompressAI BufferedRansEncoder/RansDecoder",
  "implementation": "CompressAI ANS backend used by LIC_TCM",
  "language": "C++ extension + Python/PyTorch",
  "note": "TCM's upstream model imports compressai.ans.BufferedRansEncoder and RansDecoder for bitstream coding."
}

model.text_bart_jscc_decode

Name: BART JSCC-lite text semantic decoder

Status: implemented

Inputs:

Name	Kind
symbols	channel.symbols.complex_numpy, channel.rx_symbols.complex_numpy

Optional inputs:

None.

Outputs:

Name	Kind
texts	text.batch.json

Parameters:

Name	Type	Required	Default / values	Description
cache_dir	string	no	default ``
device	string	no	default cpu
generation_max_length	integer	no	default 64
model_id	string	no	default facebook/bart-base
model_revision	string	no	default aadd2ab0ae0c8268c7c9693540e9904811f36177	Full immutable Hugging Face commit; required for remote models and matched to sender evidence.
num_beams	integer	no	default 1

Differentiability (legacy trainable_params): framework=torch, gradient=full, trainable_params=True, exportable=True

The semantic decoder is a PyTorch/Transformers module; benchmark execution uses pretrained eval mode.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	torch
dataset_capture	torch
differentiable_export	torch

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "torch",
    "implementation": "transformers_eval_artifact",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "transformers_capture_artifact",
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "transformers_decoder_module",
    "runner": "differentiable_export",
    "status": "implemented"
  }
]

model.text_bart_jscc_encode

Name: BART JSCC-lite text semantic encoder

Status: implemented

Inputs:

Name	Kind
texts	text.batch.json

Optional inputs:

None.

Outputs:

Name	Kind
symbols	channel.symbols.complex_numpy

Parameters:

Name	Type	Required	Default / values	Description
cache_dir	string	no	default ``
device	string	no	default cpu
max_length	integer	no	default 64
model_id	string	no	default facebook/bart-base
model_revision	string	no	default aadd2ab0ae0c8268c7c9693540e9904811f36177	Full immutable Hugging Face commit; required for remote models.
power_normalize	boolean	no	default True

Differentiability (legacy trainable_params): framework=torch, gradient=full, trainable_params=True, exportable=True

The semantic encoder is a PyTorch/Transformers module; benchmark execution uses pretrained eval mode.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	torch
dataset_capture	torch
differentiable_export	torch

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "torch",
    "implementation": "transformers_eval_artifact",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "transformers_capture_artifact",
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "transformers_encoder_module",
    "runner": "differentiable_export",
    "status": "implemented"
  }
]

model.text_utf8_decode

Name: Text UTF-8 payload decoder

Status: implemented

Inputs:

Name	Kind
bits	channel.payload_bits.numpy, channel.bits.numpy

Optional inputs:

None.

Outputs:

Name	Kind
texts	text.batch.json

Parameters:

Name	Type	Required	Default / values	Description
on_error	string	no	default replace values replace, fail

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

model.text_utf8_encode

Name: Text UTF-8 payload encoder

Status: implemented

Inputs:

Name	Kind
texts	text.batch.json

Optional inputs:

None.

Outputs:

Name	Kind
bits	channel.payload_bits.numpy

Parameters:

Name	Type	Required	Default / values	Description
payload_format	string	no	default json_utf8 values json_utf8

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

model.trilateration_localizer

Name: Trilateration localization baseline

Status: implemented

Inputs:

Name	Kind
problem	ai_phy.localization_problem.numpy, ai_phy.localization_observation.numpy

Optional inputs:

None.

Outputs:

Name	Kind
estimate	ai_phy.localization_estimate.numpy

Parameters:

None.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy
dataset_capture	numpy
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

model.water_filling_power_allocator

Name: Theoretical water-filling power allocator

Status: implemented

Inputs:

Name	Kind
problem	ai_phy.resource_allocation_problem.numpy

Optional inputs:

None.

Outputs:

Name	Kind
decision	ai_phy.resource_allocation_decision.numpy

Parameters:

None.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Closed-form capacity-optimal oracle for parallel Gaussian fading channels under a fixed sum-power constraint.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy
dataset_capture	numpy
differentiable_export	None

Equivalence:

{
  "reason": "Allocations must satisfy the fixed sum-power constraint and the KKT water-level solution.",
  "tolerance": {
    "atol": 1e-06,
    "rtol": 1e-05
  },
  "type": "numerical"
}

Formats:

{
  "artifact": "npz",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "closed_form_parallel_channel_water_filling",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "closed_form_parallel_channel_water_filling_labels",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

modulation

modulation.digital_modulate

Name: Digital bit-to-symbol modulator

Status: implemented

Inputs:

Name	Kind
bits	channel.payload_bits.numpy, channel.coded_bits.numpy, channel.bits.numpy

Optional inputs:

None.

Outputs:

Name	Kind
symbols	channel.symbols.complex_numpy

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
modulation	string	no	default qpsk values bpsk, qpsk, qam16

Differentiability (legacy trainable_params): framework=numpy, gradient=stop, trainable_params=False, exportable=False

Hard bit-to-constellation mapping is discrete; differentiable exports should use a differentiable modem surrogate or Sionna block.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy, cpp
dataset_capture	numpy, cpp
differentiable_export	None

Equivalence:

{
  "reason": "Hard mapper implementations should produce the same normalized constellation symbols within floating-point tolerance.",
  "tolerance": {
    "atol": 1e-07,
    "rtol": 1e-06
  },
  "type": "numerical"
}

Formats:

{
  "artifact": "npz",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "numpy_bpsk_modulator",
    "parameter_bindings": {
      "modulation": "bpsk"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "numpy_qpsk_modulator",
    "parameter_bindings": {
      "modulation": "qpsk"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "numpy_qam16_modulator",
    "parameter_bindings": {
      "modulation": "qam16"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "cpp",
    "implementation": "cpp_bpsk_modulator",
    "parameter_bindings": {
      "modulation": "bpsk"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "cpp",
    "implementation": "cpp_qpsk_modulator",
    "parameter_bindings": {
      "modulation": "qpsk"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "numpy_bpsk_modulator",
    "parameter_bindings": {
      "modulation": "bpsk"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "numpy_qpsk_modulator",
    "parameter_bindings": {
      "modulation": "qpsk"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "numpy_qam16_modulator",
    "parameter_bindings": {
      "modulation": "qam16"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "cpp",
    "implementation": "cpp_bpsk_modulator",
    "parameter_bindings": {
      "modulation": "bpsk"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "cpp",
    "implementation": "cpp_qpsk_modulator",
    "parameter_bindings": {
      "modulation": "qpsk"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

modulation.identity_modulate

Name: Identity PHY bit-to-symbol mapper

Status: implemented

Inputs:

Name	Kind
bits	channel.payload_bits.numpy, channel.coded_bits.numpy, channel.bits.numpy

Optional inputs:

None.

Outputs:

Name	Kind
symbols	channel.symbols.complex_numpy

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
label	string	no	default identity_phy_modulator

Differentiability (legacy trainable_params): framework=torch, gradient=surrogate, trainable_params=False, exportable=True

Artifact execution maps discrete uint8 bits to 0/1 complex symbols. Differentiable export may materialize this as an identity over continuous logits/symbols, but gradients through hard bits are surrogate-only.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy
dataset_capture	numpy
differentiable_export	torch

Equivalence:

{
  "reason": "Identity PHY modulation stores each bit as one finite complex symbol with real value 0 or 1.",
  "type": "exact"
}

Formats:

{
  "artifact": "npz",
  "tensor": "torch.Tensor"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "default",
    "runner": "differentiable_export",
    "status": "implemented"
  }
]

modulation.qpsk_pilot_modulate

Name: QPSK pilot-frame modulator

Status: implemented

Inputs:

Name	Kind
bits	channel.payload_bits.numpy, channel.coded_bits.numpy, channel.bits.numpy

Optional inputs:

None.

Outputs:

Name	Kind
pilot_context	channel.qpsk_pilot_context.numpy
symbols	channel.symbols.complex_numpy

Parameters:

Name	Type	Required	Default / values	Description
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
pilot_seed	integer	no	default 1701
pilot_spacing_data_symbols	integer	no	default 16
preamble_symbols	integer	no	default 16

Differentiability (legacy trainable_params): framework=numpy, gradient=stop, trainable_params=False, exportable=False

Hard QPSK mapping and discrete pilot insertion are benchmark framing operations.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy
dataset_capture	numpy
differentiable_export	None

Equivalence:

{
  "reason": "The same bits, packet layout, and pilot seed define an exact framed symbol stream.",
  "type": "exact"
}

Formats:

{
  "artifact": "npz",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

noise

noise.representation_indices

Name: Representation noise on discrete indices

Status: implemented

Inputs:

Name	Kind
indices	semantic.indices.numpy

Optional inputs:

None.

Outputs:

Name	Kind
indices	semantic.indices.numpy

Parameters:

Name	Type	Required	Default / values	Description
burst_length	integer	no	default 64
mode	string	no	default none values none, dropout, random_replace, burst
probability	number	no	default 0.0
replacement	integer	no	default 0
seed	integer	no	default 0

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

noise.representation_latents

Name: Representation-domain noise on continuous latents

Status: implemented

Inputs:

Name	Kind
latents	semantic.latents.numpy

Optional inputs:

None.

Outputs:

Name	Kind
latents	semantic.latents.numpy

Parameters:

Name	Type	Required	Default / values	Description
mode	string	no	default none values none, gaussian, dropout
probability	number	no	default 0.05
seed	integer	no	default 0
sigma	number	no	default 0.05

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Artifact-mode latent perturbation is executed as NumPy preprocessing; differentiable training export needs a torch materialization.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy
dataset_capture	numpy
differentiable_export	None

Equivalence:

{
  "reason": "Stochastic latent perturbations are equivalent by declared distribution and seed policy.",
  "type": "statistical"
}

Formats:

{
  "artifact": "npz",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

noise.source_image_perturbation

Name: Source-domain image perturbation

Status: implemented

Inputs:

Name	Kind
images	image.batch.numpy

Optional inputs:

None.

Outputs:

Name	Kind
images	image.batch.numpy

Parameters:

Name	Type	Required	Default / values	Description
blur_radius	number	no	default 1.5
mask_value	number	no	default 0.0
mode	string	no	default none values none, gaussian, blur, patch_mask, salt_pepper
patch_size	integer	no	default 32
probability	number	no	default 0.02
seed	integer	no	default 0
sigma	number	no	default 0.02

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Artifact-mode source perturbations are stochastic preprocessing blocks, not differentiable training modules.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy
dataset_capture	numpy
differentiable_export	None

Equivalence:

{
  "reason": "Stochastic perturbations are equivalent by declared distribution and seed policy.",
  "type": "statistical"
}

Formats:

{
  "artifact": "npz",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

noise.source_text_perturbation

Name: Source semantic perturbation for text

Status: implemented

Inputs:

Name	Kind
texts	text.batch.json

Optional inputs:

None.

Outputs:

Name	Kind
texts	text.batch.json

Parameters:

Name	Type	Required	Default / values	Description
mask_token	string	no	default [MASK]
mode	string	no	default none values none, drop_words, mask_words, shuffle_words
probability	number	no	default 0.0
seed	integer	no	default 0

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

source

source.ai_phy_channel_realization

Name: AI-PHY channel realization source

Status: implemented

Inputs:

None.

Optional inputs:

None.

Outputs:

Name	Kind
channel	ai_phy.channel_realization.numpy

Parameters:

Name	Type	Required	Default / values	Description
carrier_frequency_ghz	number	no	default 3.5
channel_tap_count	integer	no	default 4
delay_spread_ns	number	no	default 300.0
example_count	integer	no	default 32
mobility_kmh	number	no	default 30.0
normalize_channel	boolean	no	default True
rx_antennas	integer	no	default 1
scenario	string	no	default pilot_awgn values pilot_awgn, flat_siso, mimo_ofdm_pilot, mimo_ofdm, mimo_ofdm_3gpp_tdl
seed	integer	no	default 0
subcarrier_spacing_khz	number	no	default 30.0
subcarriers	integer	no	default 1
tdl_model	string	no	default C values A, B, C, D, E
tx_antennas	integer	no	default 1
wireless_backend	string	no	default auto values auto, numpy, sionna	Wireless implementation. auto resolves deterministically to NumPy; select sionna explicitly to require Sionna, with no runtime fallback.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Channel truth is a fixed-seed benchmark input, separate from the pilot observation model.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy, sionna
dataset_capture	numpy, sionna
differentiable_export	torch, sionna

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "npz",
  "tensor": "torch.Tensor"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "numpy_channel_realization",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "numpy_channel_realization",
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "sionna",
    "implementation": "sionna_3gpp_tdl_ofdm_channel",
    "parameter_bindings": {
      "scenario": "mimo_ofdm"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "sionna",
    "implementation": "sionna_3gpp_tdl_ofdm_channel",
    "parameter_bindings": {
      "scenario": "mimo_ofdm"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

source.ai_phy_pilot_channel

Name: AI-PHY pilot channel scenario

Status: implemented

Inputs:

None.

Optional inputs:

None.

Outputs:

Name	Kind
problem	ai_phy.channel_estimation_problem.numpy

Parameters:

Name	Type	Required	Default / values	Description
example_count	integer	no	default 32
rx_antennas	integer	no	default 1
scenario	string	no	default pilot_awgn values pilot_awgn, mimo_ofdm_pilot
seed	integer	no	default 0
snr_db	number	no	default 12.0
subcarriers	integer	no	default 1
tx_antennas	integer	no	default 1
wireless_backend	string	no	default auto values auto, numpy, sionna	Wireless implementation. auto resolves deterministically to NumPy; select sionna explicitly to require Sionna, with no runtime fallback.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Scenario generation is a benchmark data source; differentiable exports should materialize the channel model directly.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy, sionna
dataset_capture	numpy, sionna
differentiable_export	torch, sionna

Equivalence:

{
  "reason": "Pilot-observation backends must match the declared channel/noise distribution and SNR, not sample-identical noise.",
  "type": "statistical"
}

Formats:

{
  "artifact": "npz",
  "tensor": "torch.Tensor"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "numpy_pilot_awgn",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "numpy_pilot_awgn",
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "sionna",
    "implementation": "sionna_awgn_pilot_observation",
    "notes": "Selected only when wireless_backend=sionna; Sionna failures are fatal.",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "sionna",
    "implementation": "sionna_awgn_pilot_observation",
    "notes": "Selected only when wireless_backend=sionna; Sionna failures are fatal.",
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "torch_pilot_channel_scenario",
    "runner": "differentiable_export",
    "status": "implemented"
  },
  {
    "backend": "sionna",
    "implementation": "sionna_pilot_channel_scenario",
    "runner": "differentiable_export",
    "status": "implemented"
  }
]

source.ai_phy_pilot_pattern

Name: AI-PHY pilot pattern source

Status: implemented

Inputs:

None.

Optional inputs:

None.

Outputs:

Name	Kind
pilots	ai_phy.pilot_pattern.numpy

Parameters:

Name	Type	Required	Default / values	Description
pilot_spacing	integer	no	default 4
scenario	string	no	default unit values unit, comb
seed	integer	no	default 0
subcarriers	integer	no	default 1
tx_antennas	integer	no	default 1

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy
dataset_capture	numpy
differentiable_export	torch, sionna

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "npz",
  "tensor": "torch.Tensor"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "default",
    "runner": "differentiable_export",
    "status": "implemented"
  },
  {
    "backend": "sionna",
    "implementation": "default",
    "runner": "differentiable_export",
    "status": "implemented"
  }
]

source.aoa_scene

Name: Generated AoA far-field scene

Status: implemented

Inputs:

None.

Optional inputs:

None.

Outputs:

Name	Kind
scene	ai_phy.aoa_scene.numpy

Parameters:

Name	Type	Required	Default / values	Description
angle_max_deg	number	no	default 60.0	Upper bound for random broadside azimuth generation.
angle_min_deg	number	no	default -60.0	Lower bound for random broadside azimuth generation.
angle_mode	string	no	default random_uniform values random_uniform, fixed	Generate seeded random bearings or repeat one user-selected bearing.
example_count	integer	no	default 16	Number of synthetic far-field source angles generated for this run.
seed	integer	no	default 0	Controls repeatable random source-angle generation.
source_angle_deg	number	no	default 0.0	Used only in fixed mode; 0 degrees is array broadside.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy
dataset_capture	numpy
differentiable_export	torch, sionna

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "npz",
  "tensor": "torch.Tensor"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "default",
    "runner": "differentiable_export",
    "status": "implemented"
  },
  {
    "backend": "sionna",
    "implementation": "default",
    "runner": "differentiable_export",
    "status": "implemented"
  }
]

source.beamforming_scenario

Name: Beamforming scenario source

Status: implemented

Inputs:

None.

Optional inputs:

None.

Outputs:

Name	Kind
problem	ai_phy.beamforming_problem.numpy

Parameters:

Name	Type	Required	Default / values	Description
channel_model	string	no	default iid_complex_gaussian values iid_complex_gaussian, clustered_ula	Use unstructured Rayleigh fading or a three-hotspot clustered ULA distribution suitable for finite-codebook design.
example_count	integer	no	default 32
seed	integer	no	default 0
snr_db	number	no	default 10.0
tx_antennas	integer	no	default 8

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy
dataset_capture	numpy
differentiable_export	torch, sionna

Equivalence:

{
  "type": "statistical"
}

Formats:

{
  "artifact": "npz",
  "tensor": "torch.Tensor"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "synthetic_miso_channel_batch",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "synthetic_miso_channel_batch",
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "torch_miso_channel_batch",
    "runner": "differentiable_export",
    "status": "implemented"
  },
  {
    "backend": "sionna",
    "implementation": "sionna_miso_channel_batch",
    "notes": "Use Sionna PHY/RT channel materializations as this suite matures.",
    "runner": "differentiable_export",
    "status": "planned"
  }
]

source.bit_manifest

Name: Frozen bit payload manifest source

Status: implemented

Inputs:

None.

Optional inputs:

None.

Outputs:

Name	Kind
bits	channel.payload_bits.numpy

Parameters:

Name	Type	Required	Default / values	Description
item_ids	array	no	default []	Explicit ordered manifest item IDs. Mutually exclusive with selection.
manifest_path	string	yes		Local .json/.yaml/.yml noema.bit_payload_manifest path. URI schemes and symlinks are rejected.
manifest_sha256	string	yes		Expected SHA-256 of the exact manifest file bytes.
selection	string	no	default ``	Named manifest split. Mutually exclusive with item_ids.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Frozen discrete source bits are evidence artifacts and do not participate in gradient flow.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy
dataset_capture	numpy
differentiable_export	None

Equivalence:

{
  "reason": "A hash-pinned manifest and ordered selection must materialize the same canonical unpacked uint8 vector on every host.",
  "type": "exact"
}

Formats:

{
  "artifact": "npz",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

source.coco128_detection

Name: COCO128 object-detection source

Status: implemented

Inputs:

None.

Optional inputs:

None.

Outputs:

Name	Kind
detections	vision.detections.json
images	image.batch.numpy

Parameters:

Name	Type	Required	Default / values	Description
dataset_dir	string	no	default .noema/datasets/coco128
download	boolean	no	default True
image_size	integer	no	default 320
limit	integer	no	default 8
split	string	no	default train2017 values train2017

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

source.coco8_segmentation

Name: COCO8 segmentation source

Status: implemented

Inputs:

None.

Optional inputs:

None.

Outputs:

Name	Kind
images	image.batch.numpy
segmentation	vision.segmentation_mask.numpy

Parameters:

Name	Type	Required	Default / values	Description
dataset_dir	string	no	default .noema/datasets/coco8-seg
download	boolean	no	default True
image_size	integer	no	default 320
limit	integer	no	default 4
split	string	no	default val values train, val

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

source.external_classification_dataset

Name: External classification dataset source

Status: implemented

Inputs:

None.

Optional inputs:

None.

Outputs:

Name	Kind
candidate	task.predictions.json
reference	task.labels.json

Parameters:

Name	Type	Required	Default / values	Description
call_style	string	no	default dict values dict, params, none
callable	string	no	default ``
module	string	no	default ``
path	string	no	default ``

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

source.flickr8k_retrieval

Name: Flickr8k image-text retrieval source

Status: implemented

Inputs:

None.

Optional inputs:

None.

Outputs:

Name	Kind
images	image.batch.numpy
targets	retrieval.targets.json
texts	text.batch.json

Parameters:

Name	Type	Required	Default / values	Description
caption_index	integer	no	default 0
dataset_dir	string	no	default .noema/datasets/flickr8k
download	boolean	no	default True
image_size	integer	no	default 224
limit	integer	no	default 16
split	string	no	default validation values test, validation

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

source.image_dataset

Name: Real image dataset batch

Status: implemented

Inputs:

None.

Optional inputs:

None.

Outputs:

Name	Kind
images	image.batch.numpy

Parameters:

Name	Type	Required	Default / values	Description
crop_size	integer	no	default 0
dataset	string	no	default kodak	Dataset identifier. Kodak works without a manifest; every other dataset requires a hash-pinned manifest_path.
dataset_dir	string	no	default .noema/datasets/kodak
image_ids	string	no	default kodim01,kodim02,kodim03,kodim04,kodim05,kodim06,kodim07,kodim08,kodim09,kodim10,kodim11,kodim12,kodim13,kodim14,kodim15,kodim16,kodim17,kodim18,kodim19,kodim20,kodim21,kodim22,kodim23,kodim24
manifest_path	string	no	default ``	JSON/YAML noema.image_dataset_manifest. Sample paths are confined beneath its declared root.
manifest_sha256	string	no	default ``	Expected lowercase SHA-256 of manifest_path; required whenever a manifest is used.
repeat_count	integer	no	default 1
resize_shorter_side	integer	no	default 0	Optional aspect-preserving resize applied before the center crop; 0 keeps native resolution.
split	string	no	default ``	Manifest split used when image_ids is empty.
training_validation_count	integer	no	default 0	Optional external-training export control. When positive, the last N explicitly selected or manifest-split images form the validation partition; ordinary benchmark execution is unchanged.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy
dataset_capture	numpy
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

source.isac_ofdm_scenario

Name: Synthetic ISAC OFDM allocation scenario

Status: implemented

Inputs:

None.

Optional inputs:

None.

Outputs:

Name	Kind
problem	isac.ofdm_allocation_problem.numpy

Parameters:

Name	Type	Required	Default / values	Description
example_count	integer	no	default 64
seed	integer	no	default 0
sensing_weight	number	no	default 0.4
snr_db	number	no	default 10.0
subcarriers	integer	no	default 12
total_power	number	no	default 1.0

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy
dataset_capture	numpy
differentiable_export	torch

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "synthetic_frequency_selective_isac",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "synthetic_frequency_selective_isac",
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "captured_isac_allocation_contract",
    "runner": "differentiable_export",
    "status": "implemented"
  }
]

source.kodak_files

Name: Kodak image file manifest

Status: implemented

Inputs:

None.

Optional inputs:

None.

Outputs:

Name	Kind
images	image.files

Parameters:

Name	Type	Required	Default / values	Description
dataset_dir	string	no	default .noema/datasets/kodak
limit	integer	no	default 24

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

source.leo_ntn_tracking_scenario

Name: Synthetic LEO-NTN Doppler and beam-handover scenario

Status: implemented

Inputs:

None.

Optional inputs:

None.

Outputs:

Name	Kind
problem	ntn.tracking_history.numpy
truth	ntn.future_state_truth.numpy

Parameters:

Name	Type	Required	Default / values	Description
beam_count	integer	no	default 9
example_count	integer	no	default 64
history_length	integer	no	default 6
history_step_s	number	no	default 0.5
max_doppler_hz	number	no	default 48000.0
prediction_horizon_s	number	no	default 1.0
seed	integer	no	default 0
snr_db	number	no	default 15.0

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy
dataset_capture	numpy
differentiable_export	torch

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "bounded_kinematic_leo_pass",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "bounded_kinematic_leo_pass",
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "captured_future_state_supervision",
    "runner": "differentiable_export",
    "status": "implemented"
  }
]

source.local_npz_images

Name: Local NumPy image batch

Status: implemented

Inputs:

None.

Optional inputs:

None.

Outputs:

Name	Kind
images	image.batch.numpy

Parameters:

Name	Type	Required	Default / values	Description
array	string	no	default images
dtype_policy	string	no	default strict_uint8 values strict_uint8, clip_round_uint8	Reject non-uint8 input by default. clip_round_uint8 is an explicit lossy conversion that is recorded in provenance.
path	string	yes

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

source.localization_geometry

Name: Generated localization geometry

Status: implemented

Inputs:

None.

Optional inputs:

None.

Outputs:

Name	Kind
scene	ai_phy.localization_scene.numpy

Parameters:

Name	Type	Required	Default / values	Description
anchor_count	integer	no	default 4	Number of anchors generated at equal intervals around the square perimeter.
area_m	number	no	default 20.0	Side length of the generated two-dimensional localization area.
example_count	integer	no	default 16	Number of synthetic target positions generated for this run.
position_mode	string	no	default random_uniform values random_uniform, fixed	Generate seeded random targets or repeat one user-selected target position.
seed	integer	no	default 0	Controls repeatable random target placement.
target_x_m	number	no	default 10.0	Used only for fixed placement; must lie inside the configured area.
target_y_m	number	no	default 10.0	Used only for fixed placement; must lie inside the configured area.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy
dataset_capture	numpy
differentiable_export	torch, sionna

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "npz",
  "tensor": "torch.Tensor"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "default",
    "runner": "differentiable_export",
    "status": "implemented"
  },
  {
    "backend": "sionna",
    "implementation": "default",
    "runner": "differentiable_export",
    "status": "implemented"
  }
]

source.localization_sensing_scenario

Name: Localization/sensing geometry source

Status: implemented

Inputs:

None.

Optional inputs:

None.

Outputs:

Name	Kind
problem	ai_phy.localization_problem.numpy

Parameters:

Name	Type	Required	Default / values	Description
area_m	number	no	default 20.0
example_count	integer	no	default 16
range_noise_m	number	no	default 0.25
seed	integer	no	default 0
snr_db	number	no	default 20.0

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy
dataset_capture	numpy
differentiable_export	torch

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "geometric_range_measurements",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "geometric_range_measurements",
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "torch_range_geometry",
    "runner": "differentiable_export",
    "status": "implemented"
  },
  {
    "backend": "sionna",
    "implementation": "sionna_rt_scene_endpoint",
    "notes": "Future Sionna RT scenes should materialize this same range/AoA target contract.",
    "runner": "differentiable_export",
    "status": "planned"
  }
]

source.modulation_frames

Name: Balanced modulation-frame source

Status: implemented

Inputs:

None.

Optional inputs:

None.

Outputs:

Name	Kind
frames	ai_phy.modulation_frames.numpy
labels	ai_phy.modulation_labels.numpy

Parameters:

Name	Type	Required	Default / values	Description
frame_count	integer	no	default 96
seed	integer	no	default 23
symbols_per_frame	integer	no	default 128

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

The seeded symbol and label generator is benchmark data, not a learned component.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy
dataset_capture	numpy
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "npz",
  "tensor": "numpy.ndarray"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

source.near_field_xl_mimo_scenario

Name: Synthetic near-field XL-MIMO pilot scenario

Status: implemented

Inputs:

None.

Optional inputs:

None.

Outputs:

Name	Kind
problem	near_field.array_observation.numpy
truth	near_field.range_angle_truth.numpy

Parameters:

Name	Type	Required	Default / values	Description
angle_max_deg	number	no	default 55.0
angle_min_deg	number	no	default -55.0
antennas	integer	no	default 32
carrier_frequency_ghz	number	no	default 28.0
example_count	integer	no	default 64
range_max_m	number	no	default 5.0
range_min_m	number	no	default 0.5
seed	integer	no	default 0
snr_db	number	no	default 15.0

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy
dataset_capture	numpy
differentiable_export	torch

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "spherical_wave_coherent_pilot",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "spherical_wave_coherent_pilot",
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "captured_range_angle_supervision",
    "runner": "differentiable_export",
    "status": "implemented"
  }
]

source.random_bits

Name: Synthetic random bit source

Status: implemented

Inputs:

None.

Optional inputs:

None.

Outputs:

Name	Kind
bits	channel.payload_bits.numpy

Parameters:

Name	Type	Required	Default / values	Description
batch_size	integer	no	default 1
bit_count	integer	no	default 4096
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
seed	integer	no	default 0

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Synthetic discrete source bits are generated as benchmark artifacts and do not participate in gradient flow.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy
dataset_capture	numpy
differentiable_export	None

Equivalence:

{
  "reason": "Given the same seed and bit count, random-bit source materializations must produce the same canonical uint8 bit vector.",
  "type": "exact"
}

Formats:

{
  "artifact": "npz",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

source.resource_allocation_scenario

Name: OFDM subcarrier power-allocation scenario source

Status: implemented

Inputs:

None.

Optional inputs:

None.

Outputs:

Name	Kind
problem	ai_phy.resource_allocation_problem.numpy

Parameters:

Name	Type	Required	Default / values	Description
channel_tap_count	integer	no	default 4
seed	integer	no	default 0
snapshot_count	integer	no	default 32
snr_db	number	no	default 10.0
subcarrier_count	integer	no	default 16
total_power	number	no	default 1.0

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy
dataset_capture	numpy
differentiable_export	torch

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "synthetic_ofdm_frequency_response",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "synthetic_ofdm_frequency_response",
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "torch_ofdm_resource_state",
    "runner": "differentiable_export",
    "status": "implemented"
  },
  {
    "backend": "sionna",
    "implementation": "sionna_sys_resource_state",
    "runner": "differentiable_export",
    "status": "planned"
  }
]

source.retrieval_smoke

Name: Image-text retrieval smoke source

Status: implemented

Inputs:

None.

Optional inputs:

None.

Outputs:

Name	Kind
images	image.batch.numpy
targets	retrieval.targets.json
texts	text.batch.json

Parameters:

Name	Type	Required	Default / values	Description
dataset	string	no	default retrieval_smoke values retrieval_smoke
image_size	integer	no	default 224
sample_ids	string	no	default ``

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

source.semantic_artifacts_smoke

Name: Typed semantic artifact smoke source

Status: implemented

Inputs:

None.

Optional inputs:

None.

Outputs:

Name	Kind
captions_candidate	text.caption.json
captions_reference	text.caption.json
clip_embeddings	vision.embedding.clip.numpy
detections_candidate	vision.detections.json
detections_reference	vision.detections.json
importance_map	semantic.importance_map.numpy
multimodal_embeddings	multimodal.embedding.numpy
rankings	retrieval.rankings.json
scene_graph	vision.scene_graph.json
segmentation_candidate	vision.segmentation_mask.numpy
segmentation_reference	vision.segmentation_mask.numpy
semantic_map	vision.semantic_map.json
video_frames	video.frame_sequence.numpy
vqa_candidate	vqa.answers.json
vqa_reference	vqa.answers.json

Parameters:

Name	Type	Required	Default / values	Description
dataset	string	no	default semantic_artifact_smoke values semantic_artifact_smoke
embedding_dim	integer	no	default 8

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

source.task_labels_smoke

Name: Task-oriented label/prediction smoke source

Status: implemented

Inputs:

None.

Optional inputs:

None.

Outputs:

Name	Kind
candidate	task.predictions.json
reference	task.labels.json

Parameters:

Name	Type	Required	Default / values	Description
dataset	string	no	default task_smoke values task_smoke
sample_ids	string	no	default ``

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

source.text_dataset

Name: Text semantic dataset batch

Status: implemented

Inputs:

None.

Optional inputs:

None.

Outputs:

Name	Kind
texts	text.batch.json

Parameters:

Name	Type	Required	Default / values	Description
dataset	string	no	default semantic_text_smoke values semantic_text_smoke
repeat_count	integer	no	default 1
sample_ids	string	no	default weather_report,robot_instruction,medical_note,traffic_alert,sensor_summary

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

source.vqa_manifest

Name: Local VQA manifest source

Status: implemented

Inputs:

None.

Optional inputs:

None.

Outputs:

Name	Kind
answers	vqa.answers.json
images	image.batch.numpy
questions	vqa.questions.json

Parameters:

Name	Type	Required	Default / values	Description
image_root	string	no	default ``	Optional root directory for relative image paths.
image_size	integer	no	default 224
limit	integer	no	default 8
manifest_path	string	yes	default ``	JSON/JSONL file with id, image, question, and answer fields.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

source.vqa_small_hf

Name: VQA small sample source

Status: implemented

Inputs:

None.

Optional inputs:

None.

Outputs:

Name	Kind
answers	vqa.answers.json
images	image.batch.numpy
questions	vqa.questions.json

Parameters:

Name	Type	Required	Default / values	Description
dataset_dir	string	no	default .noema/datasets/vqa_small
download	boolean	no	default True
image_size	integer	no	default 224
limit	integer	no	default 8
split	string	no	default validation values test, train, validation

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

source.vqa_smoke

Name: COCO/VQA-style smoke source

Status: implemented

Inputs:

None.

Optional inputs:

None.

Outputs:

Name	Kind
answers	vqa.answers.json
detections	vision.detections.json
images	image.batch.numpy
questions	vqa.questions.json

Parameters:

Name	Type	Required	Default / values	Description
dataset	string	no	default coco_vqa_smoke values coco_vqa_smoke
sample_ids	string	no	default ``

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	python
dataset_capture	python
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "python",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "python",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

wireless

wireless.carrier_phase_impairment

Name: Residual carrier phase impairment

Status: implemented

Inputs:

Name	Kind
rx_symbols	channel.rx_symbols.complex_numpy

Optional inputs:

None.

Outputs:

Name	Kind
phase_truth	channel.carrier_phase_truth.numpy
rx_symbols	channel.rx_symbols.complex_numpy

Parameters:

Name	Type	Required	Default / values	Description
cfo_max_cycles_per_symbol	number	no	default 0.01
cfo_min_cycles_per_symbol	number	no	default -0.01
initial_phase_max_rad	number	no	default 3.141592653589793
initial_phase_min_rad	number	no	default -3.141592653589793
phase_noise_increment_std_rad	number	no	default 0.04
seed	integer	no	default 0

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

The seeded carrier process is a benchmark impairment with explicit simulation truth.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy
dataset_capture	numpy
differentiable_export	None

Equivalence:

{
  "reason": "The same packet layout, parameters, and seed define the exact carrier trajectory.",
  "type": "exact"
}

Formats:

{
  "artifact": "npz",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

wireless.channel

Name: Wireless channel over complex symbols

Status: implemented

Inputs:

Name	Kind
symbols	channel.symbols.complex_numpy

Optional inputs:

Name	Kind
channel_state	channel.ofdm_channel_state.numpy

Outputs:

Name	Kind
rx_symbols	channel.rx_symbols.complex_numpy

Parameters:

Name	Type	Required	Default / values	Description
carrier_frequency_ghz	number	no	default 3.5
channel	string	no	default awgn values awgn, flat_rayleigh, interference_awgn, mimo_flat, ofdm_tdl, ofdm_cdl, urban_micro
channel_state_mode	string	no	default none values none, explicit	explicit requires a bound channel_state artifact; none forbids one.
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
delay_spread_ns	number	no	default 100.0
fading_scope	string	no	default symbol values symbol, source_item	For flat Rayleigh fading, symbol draws an independent gain per complex symbol; source_item holds one gain constant over each source item. Source-item mode requires preserved item boundaries.
interference_sir_db	number	no	default 18.0
interferers	integer	no	default 0
mobility_kmh	number	no	default 3.0
noise_mode	string	no	default snr_at_unit_power values snr_at_unit_power, fixed_variance	Choose whether physical-channel noise is derived from unit-power SNR or set directly as variance.
noise_variance	number	no		Physical-channel noise variance used when noise mode is fixed variance.
normalize_channel	boolean	no	default True
num_ofdm_symbols	integer	no	default 14
ofdm_fft_size	integer	no	default 64
receiver_processing	string	no	default matched values matched, none	matched applies the preset’s declared receiver (perfect-CSI equalization for fading presets); none returns the raw supported receive signal.
rx_antennas	integer	no	default 1
seed	integer	no		Optional explicit operation seed. Omit it to derive the stream from the recipe master seed.
snr_db	number	no	default 12.0	Unit-power reference SNR used to derive the physical-channel noise variance.
subcarrier_spacing_khz	number	no	default 15.0
tdl_model	string	no	default A values A, B, C, D, E
tx_antennas	integer	no	default 1
wireless_backend	string	no	default auto values auto, numpy, sionna	Wireless implementation. auto resolves deterministically to the compatible NumPy materialization for ordinary sampled channels; select sionna explicitly to require Sionna. A bound explicit Sionna OFDM channel-state artifact necessarily selects Sionna.

Differentiability (legacy trainable_params): framework=torch, gradient=full, trainable_params=False, exportable=True

Native Torch and Sionna 2.x/PyTorch AWGN and flat-Rayleigh materializations preserve gradients to transmitted symbols.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy, sionna
dataset_capture	numpy, sionna
differentiable_export	torch, sionna

Equivalence:

{
  "reason": "Only materializations with the same bound channel, receiver_processing, and channel_state_mode share a statistical-equivalence obligation; random samples need not be identical.",
  "type": "statistical"
}

Formats:

{
  "artifact": "npz",
  "tensor": "torch.Tensor"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "numpy_awgn_matched",
    "parameter_bindings": {
      "channel": "awgn",
      "channel_state_mode": "none",
      "receiver_processing": "matched"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "numpy_flat_rayleigh_matched",
    "parameter_bindings": {
      "channel": "flat_rayleigh",
      "channel_state_mode": "none",
      "receiver_processing": "matched"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "numpy_interference_awgn_matched",
    "parameter_bindings": {
      "channel": "interference_awgn",
      "channel_state_mode": "none",
      "receiver_processing": "matched"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "numpy_mimo_flat_matched",
    "parameter_bindings": {
      "channel": "mimo_flat",
      "channel_state_mode": "none",
      "receiver_processing": "matched"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "numpy_ofdm_tdl_matched",
    "parameter_bindings": {
      "channel": "ofdm_tdl",
      "channel_state_mode": "none",
      "receiver_processing": "matched"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "numpy_ofdm_cdl_matched",
    "parameter_bindings": {
      "channel": "ofdm_cdl",
      "channel_state_mode": "none",
      "receiver_processing": "matched"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "numpy_urban_micro_matched",
    "parameter_bindings": {
      "channel": "urban_micro",
      "channel_state_mode": "none",
      "receiver_processing": "matched"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "numpy_awgn_unprocessed",
    "parameter_bindings": {
      "channel": "awgn",
      "channel_state_mode": "none",
      "receiver_processing": "none"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "numpy_interference_awgn_unprocessed",
    "parameter_bindings": {
      "channel": "interference_awgn",
      "channel_state_mode": "none",
      "receiver_processing": "none"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "numpy_flat_rayleigh_unprocessed",
    "parameter_bindings": {
      "channel": "flat_rayleigh",
      "channel_state_mode": "none",
      "receiver_processing": "none"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "sionna",
    "implementation": "sionna_awgn_matched_artifact",
    "parameter_bindings": {
      "channel": "awgn",
      "channel_state_mode": "none",
      "receiver_processing": "matched"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "sionna",
    "implementation": "sionna_flat_rayleigh_matched_artifact",
    "parameter_bindings": {
      "channel": "flat_rayleigh",
      "channel_state_mode": "none",
      "receiver_processing": "matched"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "sionna",
    "implementation": "sionna_mimo_flat_matched_artifact",
    "parameter_bindings": {
      "channel": "mimo_flat",
      "channel_state_mode": "none",
      "receiver_processing": "matched"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "sionna",
    "implementation": "sionna_awgn_unprocessed_artifact",
    "parameter_bindings": {
      "channel": "awgn",
      "channel_state_mode": "none",
      "receiver_processing": "none"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "sionna",
    "implementation": "sionna_flat_rayleigh_unprocessed_artifact",
    "parameter_bindings": {
      "channel": "flat_rayleigh",
      "channel_state_mode": "none",
      "receiver_processing": "none"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "sionna",
    "implementation": "sionna_explicit_tdl_ofdm_matched_artifact",
    "parameter_bindings": {
      "channel": "ofdm_tdl",
      "channel_state_mode": "explicit",
      "receiver_processing": "matched"
    },
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "numpy_awgn_matched",
    "parameter_bindings": {
      "channel": "awgn",
      "channel_state_mode": "none",
      "receiver_processing": "matched"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "numpy_flat_rayleigh_matched",
    "parameter_bindings": {
      "channel": "flat_rayleigh",
      "channel_state_mode": "none",
      "receiver_processing": "matched"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "numpy_interference_awgn_matched",
    "parameter_bindings": {
      "channel": "interference_awgn",
      "channel_state_mode": "none",
      "receiver_processing": "matched"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "numpy_mimo_flat_matched",
    "parameter_bindings": {
      "channel": "mimo_flat",
      "channel_state_mode": "none",
      "receiver_processing": "matched"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "numpy_ofdm_tdl_matched",
    "parameter_bindings": {
      "channel": "ofdm_tdl",
      "channel_state_mode": "none",
      "receiver_processing": "matched"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "numpy_ofdm_cdl_matched",
    "parameter_bindings": {
      "channel": "ofdm_cdl",
      "channel_state_mode": "none",
      "receiver_processing": "matched"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "numpy_urban_micro_matched",
    "parameter_bindings": {
      "channel": "urban_micro",
      "channel_state_mode": "none",
      "receiver_processing": "matched"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "numpy_awgn_unprocessed",
    "parameter_bindings": {
      "channel": "awgn",
      "channel_state_mode": "none",
      "receiver_processing": "none"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "numpy_interference_awgn_unprocessed",
    "parameter_bindings": {
      "channel": "interference_awgn",
      "channel_state_mode": "none",
      "receiver_processing": "none"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "numpy_flat_rayleigh_unprocessed",
    "parameter_bindings": {
      "channel": "flat_rayleigh",
      "channel_state_mode": "none",
      "receiver_processing": "none"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "sionna",
    "implementation": "sionna_awgn_matched_artifact",
    "parameter_bindings": {
      "channel": "awgn",
      "channel_state_mode": "none",
      "receiver_processing": "matched"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "sionna",
    "implementation": "sionna_flat_rayleigh_matched_artifact",
    "parameter_bindings": {
      "channel": "flat_rayleigh",
      "channel_state_mode": "none",
      "receiver_processing": "matched"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "sionna",
    "implementation": "sionna_mimo_flat_matched_artifact",
    "parameter_bindings": {
      "channel": "mimo_flat",
      "channel_state_mode": "none",
      "receiver_processing": "matched"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "sionna",
    "implementation": "sionna_awgn_unprocessed_artifact",
    "parameter_bindings": {
      "channel": "awgn",
      "channel_state_mode": "none",
      "receiver_processing": "none"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "sionna",
    "implementation": "sionna_flat_rayleigh_unprocessed_artifact",
    "parameter_bindings": {
      "channel": "flat_rayleigh",
      "channel_state_mode": "none",
      "receiver_processing": "none"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "sionna",
    "implementation": "sionna_explicit_tdl_ofdm_matched_artifact",
    "parameter_bindings": {
      "channel": "ofdm_tdl",
      "channel_state_mode": "explicit",
      "receiver_processing": "matched"
    },
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "torch_awgn_matched_module",
    "parameter_bindings": {
      "channel": "awgn",
      "channel_state_mode": "none",
      "receiver_processing": "matched",
      "wireless_backend": "auto"
    },
    "runner": "differentiable_export",
    "status": "implemented"
  },
  {
    "backend": "sionna",
    "implementation": "sionna_awgn_matched_pytorch_module",
    "parameter_bindings": {
      "channel": "awgn",
      "channel_state_mode": "none",
      "receiver_processing": "matched"
    },
    "runner": "differentiable_export",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "torch_awgn_none_module",
    "parameter_bindings": {
      "channel": "awgn",
      "channel_state_mode": "none",
      "receiver_processing": "none",
      "wireless_backend": "auto"
    },
    "runner": "differentiable_export",
    "status": "implemented"
  },
  {
    "backend": "sionna",
    "implementation": "sionna_awgn_none_pytorch_module",
    "parameter_bindings": {
      "channel": "awgn",
      "channel_state_mode": "none",
      "receiver_processing": "none"
    },
    "runner": "differentiable_export",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "torch_flat_rayleigh_matched_module",
    "parameter_bindings": {
      "channel": "flat_rayleigh",
      "channel_state_mode": "none",
      "receiver_processing": "matched",
      "wireless_backend": "auto"
    },
    "runner": "differentiable_export",
    "status": "implemented"
  },
  {
    "backend": "sionna",
    "implementation": "sionna_flat_rayleigh_matched_pytorch_module",
    "parameter_bindings": {
      "channel": "flat_rayleigh",
      "channel_state_mode": "none",
      "receiver_processing": "matched"
    },
    "runner": "differentiable_export",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "torch_flat_rayleigh_none_module",
    "parameter_bindings": {
      "channel": "flat_rayleigh",
      "channel_state_mode": "none",
      "receiver_processing": "none",
      "wireless_backend": "auto"
    },
    "runner": "differentiable_export",
    "status": "implemented"
  },
  {
    "backend": "sionna",
    "implementation": "sionna_flat_rayleigh_none_pytorch_module",
    "parameter_bindings": {
      "channel": "flat_rayleigh",
      "channel_state_mode": "none",
      "receiver_processing": "none"
    },
    "runner": "differentiable_export",
    "status": "implemented"
  }
]

wireless.digital_link

Name: Digital modulation over a simulated wireless link

Status: implemented

Inputs:

Name	Kind
bits	channel.payload_bits.numpy, channel.coded_bits.numpy, channel.bits.numpy

Optional inputs:

None.

Outputs:

Name	Kind
bits	channel.payload_bits.numpy

Parameters:

Name	Type	Required	Default / values	Description
carrier_frequency_ghz	number	no	default 3.5
channel	string	no	default awgn values awgn, flat_rayleigh, interference_awgn, mimo_flat, ofdm_tdl, ofdm_cdl, urban_micro
data_plane_backend	string	no	default auto values auto, python_numpy, cpp_native	Data-plane implementation for portable CPU kernels. auto resolves deterministically to Python / NumPy; select cpp_native explicitly to require the native extension.
interference_sir_db	number	no	default 18.0
interferers	integer	no	default 0
mobility_kmh	number	no	default 3.0
modulation	string	no	default qpsk values bpsk, qpsk, qam16
noise_mode	string	no	default snr_at_unit_power values snr_at_unit_power, fixed_variance	Choose whether physical-channel noise is derived from unit-power SNR or set directly as variance.
noise_variance	number	no		Physical-channel noise variance used when noise mode is fixed variance.
num_ofdm_symbols	integer	no	default 14
ofdm_fft_size	integer	no	default 64
receiver_processing	string	no	default matched values matched, none
rx_antennas	integer	no	default 1
seed	integer	no		Optional explicit operation seed. Omit it to derive the stream from the recipe master seed.
snr_db	number	no	default 12.0	Unit-power reference SNR used to derive the physical-channel noise variance.
subcarrier_spacing_khz	number	no	default 15.0
tx_antennas	integer	no	default 1
wireless_backend	string	no	default auto values auto, numpy, sionna	Wireless implementation. auto resolves deterministically to the compatible NumPy materialization for ordinary sampled channels; select sionna explicitly to require Sionna. A bound explicit Sionna OFDM channel-state artifact necessarily selects Sionna.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy, sionna
dataset_capture	numpy, sionna
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "operation-defined",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "sionna",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "sionna",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

wireless.miso_ofdm_csi

Name: Correlated MISO-OFDM CSI realization

Status: implemented

Inputs:

None.

Optional inputs:

None.

Outputs:

Name	Kind
csi	channel.miso_ofdm_csi.numpy

Parameters:

Name	Type	Required	Default / values	Description
carrier_frequency_ghz	number	no	default 3.5
channel_tap_count	integer	no	default 6	Tapped-delay count for the explicitly selected NumPy backend.
delay_spread_ns	number	no	default 100.0
downlink_snr_db	number	no	default 10.0
mobility_kmh	number	no	default 0.0
normalize_channel	boolean	no	default True
num_ofdm_symbols	integer	no	default 1
ofdm_fft_size	integer	no	default 32
sample_count	integer	no	default 64
seed	integer	no	default 23
subcarrier_spacing_khz	number	no	default 30.0
tdl_model	string	no	default A values A, B, C, D, E
tx_antennas	integer	no	default 8
tx_correlation_coefficient	number	no	default 0.7
wireless_backend	string	no	default sionna values sionna, numpy

Differentiability (legacy trainable_params): framework=sionna, gradient=none, trainable_params=False, exportable=False

CSI realizations are exogenous captured training data. Gradients begin at the feedback encoder input rather than flowing into random channel generation.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	sionna, numpy
dataset_capture	sionna, numpy
differentiable_export	None

Equivalence:

{
  "reason": "Backends share dimensions, average normalization, spatial correlation, delay profile, and seed semantics but do not claim identical realizations.",
  "type": "statistical"
}

Formats:

{
  "artifact": "npz",
  "tensor": "float32 RI"
}

Materializations:

[
  {
    "backend": "sionna",
    "implementation": "sionna_tdl_correlated_miso_ofdm_csi",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "sionna",
    "implementation": "sionna_tdl_correlated_miso_ofdm_csi_capture",
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "explicit_numpy_correlated_tapped_delay_fallback",
    "notes": "Selected only when wireless_backend=numpy; Sionna requests never fall back silently.",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "explicit_numpy_correlated_tapped_delay_fallback",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

wireless.modulation_awgn_observation

Name: Modulation frames with carrier uncertainty and AWGN

Status: implemented

Inputs:

Name	Kind
frames	ai_phy.modulation_frames.numpy

Optional inputs:

None.

Outputs:

Name	Kind
observation	ai_phy.modulation_iq_frames.numpy

Parameters:

Name	Type	Required	Default / values	Description
carrier_phase_max_rad	number	no	default 0.0
carrier_phase_min_rad	number	no	default 0.0
frequency_offset_max_cycles_per_symbol	number	no	default 0.0
frequency_offset_min_cycles_per_symbol	number	no	default 0.0
seed	integer	no	default 23001
snr_db_max	number	no	default 8.0
snr_db_min	number	no	default 8.0

Differentiability (legacy trainable_params): framework=torch, gradient=full, trainable_params=False, exportable=True

Complex AWGN can be reproduced by a differentiable tensor materialization.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy
dataset_capture	numpy
differentiable_export	torch

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "npz",
  "tensor": "torch.Tensor"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "default",
    "runner": "differentiable_export",
    "status": "implemented"
  }
]

wireless.ofdm_channel_state

Name: Sionna OFDM channel-state realization

Status: implemented

Inputs:

Name	Kind
symbols	channel.symbols.complex_numpy

Optional inputs:

None.

Outputs:

Name	Kind
state	channel.ofdm_channel_state.numpy
symbols	channel.symbols.complex_numpy

Parameters:

Name	Type	Required	Default / values	Description
average_power_budget	number	no	default 1.0	Authoritative normalized average transmit-power budget per OFDM subcarrier for this channel-state scenario.
capacity_multiplier	integer	no	default 1	Reserve additional OFDM channel-state capacity for allocation-aware variable-rate transport.
carrier_frequency_ghz	number	no	default 3.5
delay_spread_ns	number	no	default 100.0
mobility_kmh	number	no	default 3.0
noise_variance	number	no	default 0.1	Noise reference attached to this CSI realization; the downstream physical channel must use the same variance.
normalize_channel	boolean	no	default True
num_ofdm_symbols	integer	no	default 4
ofdm_fft_size	integer	no	default 16
seed	integer	no		Optional explicit operation seed. Omit it to derive the stream from the recipe master seed.
subcarrier_spacing_khz	number	no	default 15.0
tdl_model	string	no	default A values A, B, C, D, E

Differentiability (legacy trainable_params): framework=sionna, gradient=none, trainable_params=False, exportable=False

The artifact runner freezes a reproducible Sionna TDL realization so the allocator and channel consume identical CSI. No differentiable materialization is registered.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	sionna
dataset_capture	sionna
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "npz",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "sionna",
    "implementation": "sionna_3gpp_tdl_ofdm_frequency_response",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "sionna",
    "implementation": "sionna_3gpp_tdl_ofdm_frequency_response_capture",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

wireless.ofdm_delayed_csi

Name: Delayed/noisy OFDM transmitter CSI

Status: implemented

Inputs:

Name	Kind
state	channel.ofdm_channel_state.numpy

Optional inputs:

None.

Outputs:

Name	Kind
actual_state	channel.ofdm_channel_state.numpy
transmitter_csi	channel.ofdm_channel_state.numpy

Parameters:

Name	Type	Required	Default / values	Description
add_estimation_noise	boolean	no	default True	Add independent complex Gaussian CSI-estimation error.
allocation_ofdm_symbols	integer	no	default 24	Number of aligned old/current OFDM symbols retained per TDL block.
capture_temporal_stride	integer	no	default 1	Dataset-capture downsampling stride along each TDL trajectory. Runtime allocation and scoring retain every aligned state.
csi_estimation_snr_db	number	no	default 20.0	Mean old-channel power divided by complex CSI-estimation-error variance, in dB.
csi_history_length	integer	no	default 1	Number of consecutive causal complex-CSI snapshots exposed to a history-aware transmitter policy. The newest snapshot is still feedback_delay_ofdm_symbols older than its paired current channel state.
feedback_delay_ofdm_symbols	integer	no	default 8	CSI age in OFDM symbols. Old and current states are sliced causally from the same Sionna TDL trajectory.
seed	integer	no		Optional explicit estimation-error seed. Channel aging itself comes from the upstream TDL trajectory.

Differentiability (legacy trainable_params): framework=numpy, gradient=none, trainable_params=False, exportable=False

This operation freezes a causal pairing between old transmitter CSI and the later channel state applied by the physical link.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy
dataset_capture	numpy
differentiable_export	None

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "npz",
  "tensor": "none"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "causal_same_trajectory_csi_aging",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "causal_same_trajectory_csi_aging_capture",
    "runner": "dataset_capture",
    "status": "implemented"
  }
]

wireless.pilot_observation

Name: Noisy pilot observation

Status: implemented

Inputs:

Name	Kind
channel	ai_phy.channel_realization.numpy
pilots	ai_phy.pilot_pattern.numpy

Optional inputs:

None.

Outputs:

Name	Kind
ls_estimate	ai_phy.channel_estimation_ls_estimate.numpy
noise_variance	ai_phy.channel_noise_variance.numpy
observation	ai_phy.channel_estimation_observation.numpy
pilot_ls	ai_phy.channel_estimation_sparse_pilot_ls.numpy
pilot_mask	ai_phy.channel_estimation_pilot_mask.numpy
problem	ai_phy.channel_estimation_problem.numpy
truth	ai_phy.channel_estimation_truth.numpy

Parameters:

Name	Type	Required	Default / values	Description
seed	integer	no	default 0
snr_db	number	no	default 12.0
wireless_backend	string	no	default auto values auto, numpy, sionna	Wireless implementation. auto resolves deterministically to NumPy; select sionna explicitly to require Sionna, with no runtime fallback.

Differentiability (legacy trainable_params): framework=torch, gradient=full, trainable_params=False, exportable=True

Pilot multiplication and additive noise have a direct differentiable materialization.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy, sionna
dataset_capture	numpy, sionna
differentiable_export	torch, sionna

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "npz",
  "tensor": "torch.Tensor"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "sionna",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "sionna",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "default",
    "runner": "differentiable_export",
    "status": "implemented"
  },
  {
    "backend": "sionna",
    "implementation": "default",
    "runner": "differentiable_export",
    "status": "implemented"
  }
]

wireless.range_observation

Name: Noisy range observation

Status: implemented

Inputs:

Name	Kind
scene	ai_phy.localization_scene.numpy

Optional inputs:

None.

Outputs:

Name	Kind
observation	ai_phy.localization_observation.numpy
problem	ai_phy.localization_problem.numpy
truth	ai_phy.localization_truth.numpy

Parameters:

Name	Type	Required	Default / values	Description
nlos_bias_m	number	no	default 1.5
nlos_probability	number	no	default 0.0
range_noise_floor_m	number	no	default 0.05
seed	integer	no	default 0
snr_db	number	no	default 20.0

Differentiability (legacy trainable_params): framework=torch, gradient=full, trainable_params=False, exportable=True

Euclidean range and additive measurement noise have a direct differentiable materialization.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy
dataset_capture	numpy
differentiable_export	torch, sionna

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "npz",
  "tensor": "torch.Tensor"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "default",
    "runner": "differentiable_export",
    "status": "implemented"
  },
  {
    "backend": "sionna",
    "implementation": "default",
    "runner": "differentiable_export",
    "status": "implemented"
  }
]

wireless.ula_array_observation

Name: ULA noisy array observation

Status: implemented

Inputs:

Name	Kind
scene	ai_phy.aoa_scene.numpy

Optional inputs:

None.

Outputs:

Name	Kind
observation	ai_phy.aoa_observation.numpy
problem	ai_phy.aoa_problem.numpy
truth	ai_phy.aoa_truth.numpy

Parameters:

Name	Type	Required	Default / values	Description
antenna_count	integer	no	default 8
element_spacing_wavelengths	number	no	default 0.5
seed	integer	no	default 0
snapshot_count	integer	no	default 64
snr_db	number	no	default 15.0

Differentiability (legacy trainable_params): framework=torch, gradient=full, trainable_params=False, exportable=True

ULA steering, source mixing and additive noise have direct tensor materializations.

Training capabilities: built_in_fine_tuning=False, portable_replacement=False

Portable trained-artifact ABI: None.

Backends:

Runner	Backends
benchmark_run	numpy
dataset_capture	numpy
differentiable_export	torch, sionna

Equivalence:

{
  "reason": "No cross-backend equivalence claim declared.",
  "type": "behavioral"
}

Formats:

{
  "artifact": "npz",
  "tensor": "torch.Tensor"
}

Materializations:

[
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "benchmark_run",
    "status": "implemented"
  },
  {
    "backend": "numpy",
    "implementation": "default",
    "runner": "dataset_capture",
    "status": "implemented"
  },
  {
    "backend": "torch",
    "implementation": "default",
    "runner": "differentiable_export",
    "status": "implemented"
  },
  {
    "backend": "sionna",
    "implementation": "default",
    "runner": "differentiable_export",
    "status": "implemented"
  }
]