core/evals/deduction_serve/contract.md

Deduction-serve lane contract

What this lane scores

The production serving decider — the exact pipeline chat/deduction_surface.py::deduction_grounded_surface runs on a core chat turn: looks_like_deductive_argument (commit gate) → comprehend (reader) → the band cascade — to_deductive_logic (Band v1) → to_syllogism + the categorical decider (Band v1b, ADR-0256) → read_english_argument (Band v2-EN, ADR-0257) → read_member_argument (Band v3-MEM, ADR-0258) → read_cond_member_argument (Band v4-CM, ADR-0259) → read_verb_argument (Band v5-VP, ADR-0260) — with evaluate_entailment_with_trace (the ROBDD engine, ADR-0201/ADR-0218) deciding every band. evals/deduction_serve/runner.py::decide calls these functions directly (typed outcome, not rendered prose) — the same production decision the composer makes, without re-deriving the presentation step (generate.proof_chain.render), so this lane's pinned bytes stay stable against wording-only changes.

Splits

• v1/ — the original corpus (single-token propositional + categorical).
• v2_en/ — hand-authored REAL-English clause arguments (ADR-0257); content disjoint from the synthetic practice lexicon.
• v2_member/ — hand-authored membership/universal arguments (ADR-0258), incl. real nouns across every number-link row-type.
• v2_condmem/ — hand-authored conditional-membership arguments (ADR-0259): v2-EN's connective grammar composed over v3-MEM's singular- membership sentence reading, incl. genuine universal+connective fusion cases (a bare universal's instantiated atom unifying with a connective leaf's atom).
• v2_verb/ — hand-authored verb-predicate arguments (ADR-0260): verb universals discharged by membership facts across every closed agreement rule (+s, +es, y↔ies, irregular go/goes), transitive objects at face value, and eight typed decline shapes.

This is distinct from two existing lanes that sound similar:

• evals/deductive_logic scores the bare entail.py engine against hand-authored formula strings — it never touches the reader.
• evals/comprehension/propositional_runner.py scores reader fidelity by running the reader's projection through the independent oracle (evals.deductive_logic.oracle) as the decision procedure.

This lane is the only one that scores the production ROBDD engine (entail.py, not the oracle) end-to-end from raw text — proving the capability core chat actually serves, not just its parts in isolation.

Gold vocabulary

Four classes: entailed, refuted, unknown, declined.

declined covers every honest non-commitment: inconsistent premises (REFUSED), an out-of-band shape (categorical/syllogism, multi-word English propositions, nested negation inside an if/then clause — see "Known Band v1 boundaries" below), or a shape that doesn't even commit the turn (looks_like_deductive_argument false — not exercised by this corpus, since every committed case reads as an argument by design).

wrong=0 discipline

• wrong — the pipeline committed to a definite entailed/refuted/ unknown verdict that disagrees with gold. Must stay 0.
• declined (mismatch) — the pipeline declined on a case gold expected a definite verdict for. Not a wrong (never a confabulation), but not a pass either — the runner requires correct == n (every case's outcome class matches gold exactly, including declines matching declined gold).
• A case that gold marks declined and the pipeline also declines is correct — the lane rewards honest recognition of the boundary, not just committed accuracy.

Band-boundary history this corpus documents

Boundaries are DISCOVERED as declines, then PROMOTED to decided gold when a later band earns the shape (the corpus keeps the case, renamed …_formerly_out_of_band). Promotion does not always mean "→ entailed": the promoted gold is whatever the newly-earned band actually, honestly decides for that exact text.

• Nested negation inside if/then — a shared-reader grammar limit (not is reserved inside if/then slots); ds-v1-0006 declined until Band v2-EN decided it (promoted, ADR-0257).
• Multi-word English propositions — ds-v1-0025 declined until Band v2-EN (promoted, ADR-0257).
• Categorical/syllogism shapes — ds-v1-0023/0024/0026, decided by Band v1b since ADR-0256.
• is a membership — ds-en-0022 declined until Band v3-MEM decided it (promoted, ADR-0258).
• A bare conditional over membership clauses — ds-mem-0024 ("If Socrates is a man then Socrates is mortal. Therefore Socrates is mortal.") declined (mixed_structure_out_of_band) until Band v4-CM landed (ADR-0259), which reads it — but the antecedent is never asserted, so its honest verdict is UNKNOWN, not entailed. Promoted declined → unknown, renamed conditional_no_anchor_formerly_out_of_band — the first promotion in this corpus that does not land on entailed, because the promoted band's own correct answer for THIS text is a non-commitment, not a new capability to showcase.
• Verb predicates — refused by every band until Band v5-VP (ADR-0260) decided them (v2_verb/). NO existing case promoted: every remaining decline in the older splits uses contraction negation ("doesn't" — not in the shared tokenizer's expansion table, so v5's n't guard refuses identically) or structure v5 also guards out; the full lane's prior splits stayed byte-stable.
• Still-open declines (honest, typed): verb-phrase contraction negation (ds-en-0023/0024), ambiguous and/or scope (ds-en-0025), nested conditionals (ds-en-0026), existential quantifiers / bare plurals / definite descriptions / relative clauses / tense (ds-mem-0020…0023/0025/0026), a universal clause nested inside a connective, compound conclusions (ds-cm-0020…0026), and the verb band's own scope-outs — did not tense, multi-token subjects, prepositional/ditransitive shapes, connective×verb composition (ds-vb-0021…0028).

Reproduce

uv run python -m evals.deduction_serve.runner                              # human-facing
uv run python -m evals.deduction_serve.runner --report evals/deduction_serve/report.json  # pinned artifact

Pinned in scripts/verify_lane_shas.py as lane id deduction_serve_v1. core test --suite deductive runs tests/test_deduction_serve_lane.py, which asserts wrong == 0 and all_cases_correct is True against the committed corpus.