core/docs/adr/ADR-0257-english-clause-argument-band.md

ADR-0257 — English-clause argument band (Band v2-EN): opaque-atom propositional serving

• Status: Proposed
• Date: 2026-07-23
• Relates to: ADR-0256 (deduction-serve earned license), ADR-0201 (ROBDD keystone), ADR-0218 (proof-carrying substrate), ADR-0175/0199 (calibrated learning / arena), ADR-0253 (dual-pack serve boundary — grc/he)

1. Context

ADR-0256 shipped the first end-to-end reasoning workflow: core chat decides propositional and categorical arguments, flag-gated, behind an earned per-band SERVE license. Its two documented boundary cases were natural-English inputs:

• "If it rains then the ground is wet. It rains. Therefore the ground is wet." → declined (reserved_word_in_np);
• "If p then q. Therefore if not q then not p." (contraposition) → declined (nested negation).

Both declines came from the SHARED comprehension reader (generate/meaning_graph/reader.py), whose single-token atom floor and reserved-word guard are load-bearing for seven other lanes' wrong=0. The deductive engine itself (generate/proof_chain/entail.py) never was the limit: it is sound+complete over arbitrary propositional structure with opaque atoms. The gap was purely a reading band.

2. Decision

Add a dedicated English-clause argument reader (generate/proof_chain/english.py) used ONLY by the deduction serving path, as a fallback tier after Band v1 (propositional) and v1b (categorical):

• Whole English clauses become opaque atoms: one minted id (a0, a1, …) per distinct normalized clause; identity is normalized exact text. Clause content is never interpreted — only quoted back in the surface.
• Structure is recognized by a closed function-word grammar: if <A> then <B> (junctions allowed on both sides), [either] <A> or <B> (n-ary), top-level and (premise splits / conjunctive conclusions), and three negation forms — leading not, sentential it is not the case that, and copular <L> is|are|was|were not <R> (normalized to ~atom(<L> is <R>), contractions isn't/aren't/… expanded).
• The same verified ROBDD engine decides; deterministic templates render the verdict over the user's own clause text (generate/proof_chain/render.py::render_entailment_english).
• Four new shape-bands — en_conditional_single, en_conditional_chain, en_disjunctive, en_atomic — each earning its own SERVE license through the ADR-0199 arena at n=720/band, wrong=0 (sealed into the same ratified ledger; unlicensed ⇒ automatically hedged, per ADR-0256).

The shared reader is untouched. The composer tries v1/v1b first, so every previously-served argument is served byte-identically; this band only widens.

3. Why the opaque reading is sound (the load-bearing argument)

Propositional entailment is closed under uniform substitution of formulas for atoms. Therefore, for verdicts computed over opaque clause-atoms:

• ENTAILED, REFUTED, and inconsistent-premises verdicts remain true under ANY deeper reading of the clauses — refine a clause into internal structure, or identify two clauses we kept distinct, and the verdict stands. These are served at full strength.
• UNKNOWN ("doesn't settle") is the ONE verdict that is not substitution-safe: internal clause structure this band did not read could settle the argument. Two mitigations, both mandatory:
  1. Scoped phrasing — the UNKNOWN surface states its reading explicitly: "Reading each clause as one indivisible claim, your premises don't settle whether …". True of the reading, disclosed as such.
  2. Structural guards — clauses whose surface form ANNOUNCES structure this band cannot read refuse out of the band entirely (typed): quantifier-led categorical clauses (all/no/some/every/… — a valid syllogism must never flatten into a misleading "doesn't follow"), is a|an membership clauses (reserved for a future member_chain band), and unnormalizable negation (never, cannot, doesn't, … — a negation must never hide inside an opaque atom).

Genuinely ambiguous English (mixed top-level and+or, nested conditionals) refuses rather than guessing a scope. Honesty caps (16 premises, 24 atoms) refuse rather than truncate.

4. What the earned licenses certify

Per ADR-0256, a band's license certifies READER fidelity per shape — hence the en_ namespace: same engine, different reader, so the English reader earns its own record. The arena corpus is synthetic copular clauses ON PURPOSE (the reader is content-blind; the license certifies structural fidelity); the eval lane (evals/deduction_serve/v2_en/, 26 hand-authored REAL-English cases, content disjoint from the synthetic lexicon) keeps that claim honest against natural prose — the synthetic-corpus-overfit lesson applied in design.

Corpus gold is authored by construction and cross-checked against the independent truth-table oracle with NO reader in the loop (the template's intended logical form is checked directly) — stronger independence than the v1 path, INV-25 preserved.

5. Tri-language extension contract (grc / he)

CORE's three foundational languages are a capability asset precisely where English is structurally weak, and this band is built to receive them:

• What is language-neutral (reused as-is): the opaque-atom table, the substitution-soundness argument (§3), the caps, the refusal discipline, the license machinery, the engine. None of it knows English.
• What is language-specific (the thin layer a sibling band replaces): the structural lexicon (if/then/or/and/not/either/therefore, copulas, contractions, quantifier leads) and the clause-order conventions.
• Honest wrinkle, recorded so we don't build a hollow abstraction: a pure lexicon swap does NOT read real Greek — ἄρα is postpositive (second position), and case-driven word order changes clause segmentation. A grc_* / he_* band is a sibling reader module + its own earned licenses, entering through pack ratification (ADR-0253: draft he/grc trees are not serve imports). No premature parameterization is introduced now.
• The trigger (per the ratified pack-expansion doctrine): when en-band refusal telemetry shifts from mechanism (ambiguous_and_or — English genuinely ambiguous) to coverage, that is the evidence that the morphologically explicit languages pay — they read cleanly exactly where English refuses.

6. Scope-outs (deliberate)

1. member_chain band (membership + quantified premises — "Socrates is a man. All men are mortal. Therefore Socrates is mortal."): guarded out (membership_shape_out_of_band), reserved as the next band; needs a per-individual propositional lowering (same pattern as generate/proof_chain/categorical.py).
2. Verb-phrase negation de-conjugation ("doesn't start" → "starts"): refuses today (internal_negation_unread); English morphology work, or the tri-language path (§5).
3. Proof-step recap in the surface: unchanged from ADR-0256 scope-out.
4. Flag posture unchanged: the band rides deduction_serving_enabled (default off); no new flag.

7. Verification

• uv run core test --suite deductive -q — 83 passed (reader contract, surfaces, license, both lane splits, e2e through ChatRuntime).
• Practice arena: 9 bands × 720 = 6,480 cases, wrong=0, every band reliability 0.99087 ≥ θ_SERVE=0.99 → SERVE licensed; ledger re-sealed (chat/data/deduction_serve_ledger.json, self-verifying content_sha256).
• Eval lane: v1 28/28 (the two former boundary declines now decided — entailed), v2_en 26/26, wrong=0; report re-pinned (scripts/verify_lane_shas.py::deduction_serve_v1).