Skip to content

Evidence and Test Matrix

Meridian Verity Group edited this page Jul 19, 2026 · 1 revision

Evidence and Test Matrix

Interpretation rule

The table below reports values published for v2.2.6-public-eval. A green count demonstrates that the published harness observed the expected results in its measured source tree. It is not, by itself, production certification, formal conformance, proof of non-bypassability, or independent validation.

Evidence family Command Published v2.2.6 result What it helps establish
Static source manifest python verify_manifest.py PASS Tracked public source agrees with the published manifest.
Public packet validation python tools/validate_public_eval_packet.py PASS Required reviewer-facing files are present and internally consistent.
Release lineage python tools/check_release_lineage.py PASS Active version and release tuple point to the intended release.
IETF release pointers python tools/check_ietf126_release_pointers.py PASS / 0 findings Reviewer links target the current public-evaluation tag.
Selected IETF packet python ietf126/run_review_packet.py 20 / 20 PASS Selected field, vector, and behavior checks match expected outcomes.
Independent recomputation python ietf126/independent_recompute.py 17 / 17 PASS Canonical bytes and digests are recomputed without importing the main verifier package.
Independent crypto python ietf126/independent_crypto_verify.py 19 / 19 PASS Receipt, revocation, and carrier signatures are checked through an independent path.
Public evaluation vectors python run_vectors.py 76 / 76 PASS Positive and fail-closed negative vectors match expected decisions.
Strict pytest PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 PYTHONWARNINGS=error python -m pytest -q -p no:cacheprovider 323 / 323 PASS Automated unit and integration expectations pass with plugins disabled and warnings fatal.
Full package coverage make coverage 99.33% statements / 98.44% branches Broad execution coverage across orprg_eval; not a security proof.
Security-critical coverage make coverage 100.00% lines / 99.705% branches High measured coverage across the explicitly selected security-critical module set.
Verifier branches make coverage 100.00% Every measured verifier branch executed in the published suite.
Strict-schema branches make coverage 100.00% Every measured strict-schema branch executed in the published suite.
Full release preflight make qa-full PASS Aggregated evaluation, metadata, release, and integrity gates pass.
Deterministic build python tools/build_release_asset.py ... twice plus byte comparison exact-byte comparison required The four release assets are reproducible in the locked environment.

Evidence hierarchy

Use evidence in this order:

  1. Immutable bytes — ZIP and sidecars downloaded from the release.
  2. Cryptographic bindings — checksum, manifest, provenance, and source tag.
  3. Re-execution — commands run in a recorded clean environment.
  4. Independent paths — recomputation and crypto verification that do not simply call the main verifier.
  5. Test reports and coverage — useful breadth indicators, not substitutes for threat analysis.
  6. Narrative claims — README, release notes, and wiki summaries.

What coverage does not prove

High line or branch coverage does not prove that:

  • every effect-relevant field was included in canonicalization;
  • test assertions are correct;
  • cryptographic trust configuration is appropriate;
  • all production paths traverse the gate;
  • replay state is durable under every deployment failure mode;
  • policy is correct or non-ambiguous; or
  • an attacker cannot find an unmodeled state or integration path.

Coverage should be paired with negative vectors, independent recomputation, mutation testing, structured fuzzing, external implementations, and deployment-level bypass analysis.

Independent reviewer record

A credible external reproduction should publish:

  • release tag and commit;
  • downloaded artifact hashes;
  • operating system, architecture, Python, pip, and dependency mode;
  • exact commands;
  • full pass/fail counts;
  • deviations from documented setup;
  • any generated file differences; and
  • a statement that no production or sensitive data was used.

Use Contributing and Feedback for public-safe findings.

Clone this wiki locally