Noema Documentation#

Noema launch hero stating Make the comparison traceable, with the bounded experimental receiver headline and its scientific-status disclosure

Noema is an experiment-contract and evidence system for learned-communication comparisons. It keeps the protocol, execution plan, resource accounting, returned model, terminal outcome, and retained evidence attached to the same verifiable object.

Noema’s verifier establishes local, profile-scoped consistency. It does not establish scientific fairness, standards conformance, authenticity, or independent reproduction.

Start with the failure#

Break the comparison is the shortest introduction to Noema. Begin with a valid paired result, then change the condition, aggregation, metric, or comparator role. The arithmetic remains precise while the claim loses the contract that made it interpretable.

Watch the complete workflow#

Contract export, dataset capture, external training, returned-model validation, and comparison with uncompensated and calibrated-oracle baselines. Watch on YouTube

Choose your route#

Run an experiment Install from source, instantiate a recipe, and retain a verifiable result. Train your method Export a typed training contract and return a portable trained artifact. Study examples Open the physical-layer and end-to-end communication demonstrations. Audit evidence Understand manifests, traceability profiles, and verification boundaries. Bring your own model Connect an encoder, metric, dataset, or portable inference runtime. Prepare a publication Separate completed evidence, release controls, and scientific readiness.

How Noema works#

Six-step Noema workflow from opening a scenario and choosing trainable blocks through contract export, external model training, recipe binding, and verified benchmarking — Noema’s contract-first model-development loop. Open the diagram for a full-size view or download the editable draw.io source.#

An executable experiment contract links a schema-validated protocol to a concrete execution plan, declared rate and channel-use accounting, optional returned-model interfaces, terminal outcomes, and retained result and plot evidence. Noema runs typed benchmark and capture recipes, exports architecture-neutral training-return contracts, and checks the identities and relations covered by the selected traceability profile.

Evidence, with its boundaries#

Log-scale pre-decoder BER versus SNR for uncompensated QPSK, a calibrated diagnostic oracle, and a learned receiver with an observed min-max band over paired seeds — The selected launch demonstration. The learned band is the observed minimum and maximum over three paired held-out seeds—not a confidence interval—and the calibrated oracle is a diagnostic reference. Open the chart for a full-size view.#

The demo, chart, and tables read their values and warnings from the same generated evidence file, so their claims cannot silently drift apart. Open the complete F0–F5 visual and T0–T2 table set, inspect the canonical launch evidence, or browse the static result explorer. This is completed experimental evidence, not a publication-ready canonical benchmark.

Browse the documentation#

Use the navigation menu for the full documentation tree. New readers can start with the tutorials, demonstrations, architecture, or generated reference. For the source-checkout quickstart, see the project README.