[hippoley]

Reality signal

This feature exists because an external human reviewer asked for it twice.

In Reality Probe #13, sylvesterkaczmarek said the useful review shape is a compact claim/evidence matrix with:

• exact review claim / failure mode;
• exact tests;
• BASE and HEAD;
• oracle alignment;
• explicit WITNESSED / CONTRADICTED / UNPROVEN boundaries;
• short assertion/failure excerpts.

After seeing the manual #731 matrix, the reviewer confirmed the shape and asked for two changes before automation:

1. fixed oracle vocabulary: UNVERIFIED / ALIGNED / CONTRADICTED;
2. a red→green row with no oracle check must be scoped as WITNESSED (submitted judge), not bare product-level correctness.

In Reality Probe #16, the same reviewer independently confirmed that oracle alignment should remain a separate evidence axis and that soft wording such as SUPPORTED should not silently upgrade evidence.

Smallest productization

Add:

counterproof claim-matrix claims.yml

The command does not infer claims and does not choose an authoritative oracle.

A human/upstream review system supplies an explicit manifest. CounterProof only:

• validates mechanical evidence vocabulary;
• requires exact tests + BASE/HEAD for witnessed/not-witnessed rows;
• requires an explicit oracle probe for ALIGNED/CONTRADICTED rows;
• derives a conservative overall claim;
• emits Markdown + JSON.

Derived semantics

WITNESSED + ALIGNED      -> PROVEN
anything + CONTRADICTED  -> CONTRADICTED
WITNESSED + UNVERIFIED   -> WITNESSED (submitted judge)
otherwise                -> UNPROVEN

Reality acceptance fixtures

Two public cases are encoded as examples and tests:

• openai/codex-plugin-cc#731
• anthropics/claude-code#89404

The test suite explicitly rejects ambiguous vocabulary such as SUPPORTED.

Non-goals

• no LLM claim extraction;
• no automatic oracle discovery;
• no merge recommendation;
• no replacement for Regression Witness or human review.

This is deliberately just the artifact shape a real reviewer said would reduce review uncertainty.