Benchmark Submissions#
This document defines the local submission rules for Noema benchmark results.
Submission route#
Validate the complete bundle locally as described below.
Open a proposal at M0574F4/noema-lab#new with the benchmark ID, comparison division, bundle size, and a digest. Do not attach private data or security-sensitive details.
Submit small metadata/fixture changes through a pull request. For large result bundles, wait for the acting maintainer to approve an immutable artifact location before upload.
The current repository does not operate a private leaderboard or automatic intake service.
Allowed Modifications#
Closed division submissions may change only the method contribution:
codec, semantic encoder, receiver, channel model, channel code, modulation block, metric, or dataset adapter when the benchmark allows that contribution type;
runtime/backend choices declared in the recipe and manifest;
method-specific hyperparameters declared in the recipe.
Closed division submissions must not change:
canonical benchmark dataset/split/sample IDs;
metric definitions or reductions;
rate/channel fixed points;
reference data or ground truth;
benchmark pack ID/version;
seed policy, except where the protocol explicitly defines a seed grid.
Open division submissions may change more of the system, but must declare every difference in the submission metadata and should not be compared directly with closed-division results.
Required Metadata#
Submissions must include:
method name and authors;
comparison division:
closedoropen;adapter manifest path/name/version when an adapter is used;
benchmark ID and version;
result bundle path;
recipe SHA-256;
hardware, runtime, software dependencies, and optional accelerator details;
seed policy and concrete seeds;
any external data, checkpoint, or pretrained artifact identifiers.
Validate Locally#
uv run noema submission validate path/to/submission.yaml
The validator requires a normalized relative result_bundle path confined to the submission
directory and rejects traversal, absolute paths, and symlinks. It invokes the full benchmark verifier,
including result-local run evidence, protocol identity, artifacts, reports, plots, resource admission,
and training-lineage checks. It also cross-checks every recipe digest, concrete master seed, recorded
seed-policy digest, and the comparison division frozen in benchmark.json.
The default successful verdict is verified_internal_consistency_untrusted_producer. That verdict is
appropriate for an honest producer and accidental-drift detection; it does not authenticate hostile
input. API callers can pass an independently obtained trusted_bundle_root_sha256 to receive
verified_pinned_bundle_integrity after the exact bundle inventory matches. A pin proves byte identity,
not author identity or scientific truth; signatures and producer identity remain release-policy
responsibilities outside this local validator.
For multi-recipe results, reproducibility.recipe_sha256s must map every benchmark recipe ID to the
recorded run recipe digest. reproducibility.seed_policy_sha256s must list the distinct SHA-256
identities of the full recorded seed-policy objects. A single-recipe submission may put the one
seed-policy digest directly in seed_policy; the human-readable field remains available when the
digest list is supplied. The frozen benchmark must declare either
metadata.comparison_division or metadata.allowed_comparison_divisions.
Minimal Submission YAML#
schema_version: 1
comparison_division: closed
result_bundle: .noema/benchmarks/20260708T000000Z_benchmark_v1.image_reconstruction.kodak
method:
name: My Semantic Codec
authors:
- A. Researcher
adapter:
manifest: adapters/my_codec/noema_adapter.yaml
name: my_codec
version: 0.1.0
benchmark:
id: benchmark_v1.image_reconstruction.kodak
version: "1.0.0"
runtime:
hardware: CPU
software: noema 0.2.0.dev0
dependencies:
- numpy
reproducibility:
recipe_sha256: "<sha256 from manifest>"
seed_policy: noema deterministic step-derived seeds
seed_policy_sha256s:
- "<sha256 of the complete recorded seed_policy object>"
seeds:
- 23