core/docs/research/english-argument-band-2026-07-23.md

English-clause argument band (Band v2-EN) — natural-language deduction serving, 2026-07-23

Base: main @ 1a6ccaf9. Branch: feat/english-argument-band. ADR: docs/adr/ADR-0257-english-clause-argument-band.md.

The gain, stated as transcripts

Everything below declined yesterday. It is decided today — same sound engine, wrong=0 discipline intact, licenses earned not granted:

> If it rains then the ground is wet. It rains. Therefore the ground is wet.
Given: if it rains then the ground is wet; it rains. Your premises entail: the ground is wet.

> If the alarm is armed then the light is on. The light is not on. Therefore the alarm is not armed.
Given: if the alarm is armed then the light is on; the light is not on. Your premises entail: the alarm is not armed.

> Either the door is open or the window is open. The door isn't open. Therefore the window is open.
Given: either the door is open or the window is open; the door is not open. Your premises entail: the window is open.

> If the pump runs then the tank fills. If the valve opens then the tank fills. The pump runs or the valve opens. Therefore the tank fills.
Given: …. Your premises entail: the tank fills.        (constructive dilemma)

> If p then q. Therefore if not q then not p.
Given: if p then q. Your premises entail: if not q then not p.   (contraposition — the other former boundary case)

The two texts ADR-0256's completion doc recorded as "the future acceptance tests" (out_of_band_multiword_conditional, out_of_band_nested_negation) are both now decided — their eval-lane gold updated from declined to entailed accordingly.

What was built

Piece	File	Role
English argument reader	generate/proof_chain/english.py (new)	Closed function-word grammar over OPAQUE clause-atoms; minted atom ids; typed refusals; caps
English renderer	generate/proof_chain/render.py::render_entailment_english	Verdicts over the user's own clauses; UNKNOWN explicitly scoped to the opaque reading
Band constants	generate/proof_chain/shape.py	en_conditional_single/chain, en_disjunctive, en_atomic; ALL_SHAPE_BANDS
Composer fallback tier	chat/deduction_surface.py::_english_band_surface	Tried strictly AFTER v1/v1b — monotone widening, byte-identical on everything previously served
Arena extension	evals/deduction_serve/practice/gold.py	4 en-bands × 720 synthetic copular-clause cases; gold by construction, oracle-checked with NO reader in the loop
Sealed ledger (9 bands)	chat/data/deduction_serve_ledger.json	All 9 bands wrong=0 at n=720 → reliability 0.99087 ≥ θ_SERVE=0.99 → SERVE earned
Real-English eval lane	evals/deduction_serve/v2_en/cases.jsonl (26 cases)	Hand-authored natural prose, content disjoint from the synthetic lexicon — anti-overfit guard
Tests	tests/test_english_argument_reader.py (new, 33 asserts across 20 tests) + extensions	Reader contract, guards, surfaces, license hedging, e2e through ChatRuntime

Why it is sound (the one theorem the design leans on)

Propositional entailment is closed under uniform substitution. So ENTAILED / REFUTED / inconsistent verdicts over opaque clause-atoms stay true under ANY deeper reading of the clauses — those serve at full strength. UNKNOWN is the one non-substitution-safe verdict, so it (a) renders with an explicitly scoped phrasing ("Reading each clause as one indivisible claim…"), and (b) is protected by structural guards: quantifier-led categorical clauses, is a/an membership clauses, and unnormalizable negation all refuse OUT of the opaque band (typed reasons), so a real syllogism can never be flattened into a misleading "doesn't follow". Ambiguous scope (and+or mixed, nested conditionals) refuses rather than guesses.

Earned, not granted

The en_* bands ride the exact ADR-0256 license machinery: absent from the ledger ⇒ hedged automatically (tested); present with wrong=0 at volume ⇒ authoritative. The arena run that sealed them: 9 bands × 720 = 6,480 cases, wrong=0 everywhere, in ~1.8s. Strip the ledger and every answer above still comes back — sound but DISCLOSED. That remains the difference between a flagged capability and an earned one.

Tri-language posture (per direction, recorded in ADR §5)

The atom table, soundness argument, caps, refusal discipline, license machinery, and engine are language-neutral; only the structural lexicon + clause-order conventions are English. A grc_*/he_* sibling band is a reader module + its own earned licenses through pack ratification (ADR-0253 boundary respected) — NOT a premature lexicon parameter (ἄρα is postpositive; a lexicon swap alone cannot read real Greek; recorded so we don't build a hollow abstraction). The en-band refusal telemetry (internal_negation_unread, ambiguous_and_or) is the instrumented trigger: it marks exactly where English is the bottleneck and the explicit-morphology languages would read clean.

Honest scope-outs

1. member_chain band ("Socrates is a man. All men are mortal. Therefore Socrates is mortal.") — guarded out with a typed reason; the designed next band (per-individual propositional lowering, same pattern as categorical.py).
2. Verb-phrase negation ("doesn't start") — refuses; de-conjugation is real morphology work (or the tri-language path).
3. Proof-step recap — verdicts only, unchanged scope-out.
4. Flag — still deduction_serving_enabled=False default; no new flag; flip remains ratification.

[Verification]

uv run core test --suite deductive -q     # 83 passed
uv run core test --suite smoke -q          # (pre-push gate — run in worktree)
uv run core test --suite cognition -q      # (regression: shared reader untouched)
uv run python -m evals.deduction_serve.practice.runner        # 9 bands SERVE, wrong=0
uv run python -m evals.deduction_serve.runner                 # v1 28/28, v2_en 26/26
# ledger self-verifies content_sha256; lane report re-pinned (verify_lane_shas.py)