core/evals/frame_verdict_text_cwa

frame_verdict_text_cwa — text closed-world evaluation lane (B4 PR-2)

A measure-only lane over the closed-world FrameVerdict text evaluator (generate/frame_verdict/evaluate.py, ADR-0222). It proves the sealed type can evaluate text/propositional closed frames safely — including a sound negative conclusion (entailed_false) — without touching the open-world determine() path or runtime serving.

What it checks

For each v1/cases.jsonl case (a ClosedFrame + query + hand-authored gold):

• the production evaluate_frame_verdict verdict == gold (wrong = 0);
• the gold is independent — oracle.py::oracle_frame_verdict re-derives it by its OWN recursive-descent parser + brute-force truth-table enumeration, disjoint from the ROBDD (imports no engine module; never calls evaluate_frame_verdict / determine / response_governance). Engine and oracle must agree on every case (INV-25/27 discipline);
• every entailed_false carries an admissible proof (proof_chain.entail / REFUTED / ROBDD_REFUTATION, non-empty proof_sha256); a generic FALSIFIED is impossible (the type's __post_init__ invariant);
• absence / OPEN / undeclared-closure are never entailed_false — they refuse (UNDETERMINED / SCOPE_BOUNDARY).

Input contract (PR-2)

propositions and query are propositional-formula strings in the proof_chain.entail grammar (atoms, ~, &, ->, parens; ~ > & > ->, -> right-assoc). There is no prose lowering — a prose front-end is out of scope. Malformed / out-of-regime input refuses with SCOPE_BOUNDARY.

Gold categories

ENTAILED_TRUE · ENTAILED_FALSE · UNDETERMINED · CONTRADICTION · SCOPE_BOUNDARY.

Independence (honest scope note)

The disjoint solver is the wrong=0-critical part: the truth-table oracle decides entailment independently of the ROBDD. The frame-gating contract (OPEN / undeclared / non-TEXT => SCOPE_BOUNDARY) is shared by both — it is the frame contract, not a solver, so both apply it; independence is in the solving.

Capability-index: deliberately NOT registered

This lane is not added to evals/capability_index and does not change the capability baseline/digest. The whole closed-world feature is off-serving (sealed type + INV-31 firewall; default-dark even after the governance slice), so registering a closed-world serving capability domain would be premature. ADR-0222 §9 anticipates a future ProofWriter-CWA capability lane registering once closed-world becomes a ratified serving capability; until then this stays a measure-only correctness lane (the posture of the #779 ProofWriter-OWA floor).

Run

python -m pytest tests/test_frame_verdict_text_cwa_lane.py
python -m evals.frame_verdict_text_cwa.score   # prints the report; exit 1 on wrong>0