core/docs/assessment/10-layer-cards/M3-comprehension-reasoning.md

M3 — Comprehension & Reasoning

Kind: layer · Parent: CORE · Assessor: Opus 5 (Phase 2) Re-verified: 39331dbc (2026-07-28) — see the arc note below. Verified at: 8927c563 (2026-07-27) Liveness: partial-wiring-debt · Fitness: strained (with one superseded-in-place sub-region) · Topology role: runtime boundary

The wiring that thinks. Per the governing mental model — the field is electricity, the intelligence is in the wiring — M3 is where CORE's intelligence actually lives. It takes what M2 admitted and what M1 knows, and produces a decided proposition: what is the case, what follows, what cannot be determined. Its governing competence standard is ADR-0252: comprehension must grasp deep relational structure that subsumes a family of problems, not surface features of one.

Telos stages: comprehend, think/reason (primary); recall (in-turn) Macro role: Turns admitted input into a decided, evidence-bearing proposition graph — or a typed refusal.

Arc note — 2026-07-28 (39331dbc). Track A ran ADR-0252 §5 to a NO-GO (docs/research/sme-experiment-verdict-797ebad5.md), settling the paradigm question this layer's successor path depended on. New measurement G-21: the math reader returns a selected graph for 5 of 500 holdout_dev/v1 cases (1.0%), all one skeleton — a sharper statement of the comprehension frontier than the 19-construction count. H-8e corrected an inductive-closure docstring that described a cycle check the code does not (and must not) perform.

---

What it is / What it does

M3 spans recognition (recognition/, teaching-derived structural recognizers via anti-unification), the cognition pipeline (core/cognition/, 13 modules, pipeline.py at 1381 lines), the comprehension attempt/router (core/comprehension_attempt/: classify.py, failure_family.py, router.py, proposal.py), the determine phase (generate/determine/: determine.py, consolidate.py, estimate.py, estimation_license.py, render.py), the realize phase (generate/realize/, generate/realizer.py), the deduction flagship (generate/proof_chain/ — ROBDD tautology engine under six ratified bands), curriculum-grounded reasoning, the GSM8K math reader (demoted to diagnostic), and the field-wedge research zone (a recorded negative result).

Operationally, the live path enters M3 after ingest: intent classification → PropositionGraph construction → determination or entailment → ArticulationTarget. Deduction serving is ratified ON (deduction_serving_enabled=True, core/config.py:397); curriculum serving is OFF (curriculum_serving_enabled=False, :412), as are ask_serving_enabled and verified_serving_enabled. The deduction engine is generate/proof_chain/entail.py — canonicalize (P1 & … & Pn) → Q, ask is_tautology over a ROBDD. It is a decision procedure, not a derivation: EntailmentTrace carries an outcome, a reason, and five canonical BDD node keys (opaque hashes). There are no intermediate proof steps in the object, which is why "the renderer drops the proof chain" was previously diagnosed at the wrong layer — nothing is dropped because nothing exists to drop. Multi-step articulation is engine work.

---

Contract

• Inputs: injected FieldState, tokenized/OOV-policed text, mounted packs and vocabulary (M1), session context.
• Outputs: PropositionGraph (now carrying GraphNode.negated, ADR-0265), ArticulationTarget, EntailmentTrace / Determined / Undetermined, typed refusals, recognizer outcomes, ContractAssessment.
• Invariants:
  • INV-30 — open-world determine() constructs only Determined(answer=True) or refuses; never asserts False — pin: tests/test_architectural_invariants.py — suite: present in the test tree; suite membership unverified at this SHA (flagged).
  • INV-31 — closed-world FrameVerdict cannot reach the open-world runtime (transitive import containment + single construction allowlist + typed refusal) — pin: tests/test_architectural_invariants.py.
  • INV-34 — cognition-pipeline failures are typed, never silent; a None ContractAssessment is itself a violation; unresolvable referents refuse rather than fill; no PASSTHROUGH in the intent ratifier — pin: tests/test_linguistic_governance_phases.py.
  • Kernel no-new-legacy — new derivation must consume ProblemFrame/KernelFacts; new raw-prose regex requires an explicit LEGACY_EXCEPTION — pin: tests/test_kernel_no_new_legacy_derivation_surfaces.py.

---

Design vs build

• Design: ADR-0252 (Accepted, governing) — the problem-solving paradigm and its §4 conformance bar; ADR-0251 (halt bespoke per-case regex work; the prohibition governing every math-reader increment); ADR-0249/0250 (reader→Hamiltonian compiler); ADR-0256–0261 (six deduction bands); ADR-0262/0264 (curriculum); ADR-0265 (negation in the proposition graph); ADR-0243 (wave-field lifecycle); ADR-0142 (epistemic taxonomy).
• Build: partial-wiring-debt.
• Evidence:
  • Deduction serving decides real arguments end-to-end, wrong=0 across all splits — lane — evals/deduction_serve/report.json, pinned SHA 0b461a5a… in CLAIMS.md — would-fail-if-absent: yes.
  • Propositional entailment scored against an independent truth-table oracle, 716/716 correct, wrong=0, refused=0 — lane — evals/deductive_logic/report.json, pinned 97a23094… — would-fail-if-absent: yes.
  • deductive suite exists and carries 20 test files — code-read — core/cli_test.py — would-fail-if-absent: yes.
  • 25 sealed bands / 18,000 cases / wrong=0; capability index breadth 13, wrong_total=0 — measurement — chat/data/deduction_serve_ledger.json, evals/capability_index/baseline.json (confirmed mechanically 2026-07-25).
  • Structure-mapping (the ADR-0252 §6 correction) — not built. conformal_procrustes exists in core/physics/dynamic_manifold.py but is off-serving and unwired to comprehension. would-fail-if-absent: no.

---

Capacity

• Designed: comprehension that grasps deep relational structure, one canonical structure subsuming a family (generalization ratio > 1).
• Measured — the two-grammars result, and it is the sharpest number in the assessment. Reader and writer inventories measured against each other (PR #138, c69f9948): the writer emits 1739 constructions; the reader comprehends 19; the overlap is 6, all set-theoretic. The reader fabricates on 22 more — every dog is a mammal → member(every_dog, mammal); Given: furthermore; p implies q; p. reaching served output with furthermore recited back as a premise. These are measured and pinned, held for ADR + ratification — recorded here, not re-discovered and not fixed. Prior prose claiming the reader "reads" its arguments was falsified: g_args_rate was 0.0 throughout.
• Ceilings: curriculum bands are capped at ≤16 entailed cases by the 16-premise compilation cap (ADR-0264 §4.1), so no curriculum band can earn SERVE until query-scoping lands — an engineering blocker, not a content one. Deduction's ROBDD decides but cannot narrate multi-step derivations. The GSM8K sealed holdout (1,319 cases) has never been opened against a parser with sufficient coverage.

---

Dependencies & provenance

reads → M1 (packs, vocabulary, vault recall), M0 (field, CGA); receives ← M2; feeds → M4 (articulation and serve seam), M5 (refusals and comprehension failures become proposal candidates); verified-by → MV; governed-by → MG.

---

Stage coverage

Stage	Verdict	Evidence
comprehend	covered, narrowly	Deduction bands decide real arguments with wrong=0; but the reader spans 19 constructions against a 1739-construction writer and fabricates on 22 — coverage is real and thin
think/reason	covered	ROBDD entailment, 716/716 against an independent oracle; idle consolidation climbs to deductive closure under proof-gating
recall (in-turn)	covered	Exact CGA recall via M1

Zone roster: L4-recognition, L5-cognition, comprehend-organ ⚑, determine-phase ⚑, realize-phase ⚑, reasoning-deductive, gsm8k-math, field-wedge-research.

⚑ Zero-subsystem zones — honest scoping. Three of the four unmapped zones live here and all three sit on the serving path. What is known at this SHA: comprehend-organ corresponds to core/comprehension_attempt/ (6 modules: classify, failure_family, router, proposal, model); determine-phase to generate/determine/ (8 modules incl. consolidate, estimate, estimation_license, render); realize-phase to generate/realize/ + generate/realizer.py (306 lines) + generate/realizer_guard.py. What is unmapped: their internal decomposition, per-component liveness, and the boundary between comprehension_attempt and L5-cognition. These are Phase 3's first descent targets.

Rollup note: weakest-link rollup; not a completion rate. M3's deduction region is genuinely strong (ratified, evidenced, wrong=0); its comprehension region is measurably narrow. A single layer label cannot express that split, which is why the stage table above separates them.

---

Judgment

Fitness: strained, with the 18 derivation organs superseded-in-place.

The strain is precisely what ADR-0252 named: an expert substrate driven by a novice reader. The ADR is ratified, its diagnosis stands, its §6 correction is designed — and its §5 acceptance gate has never returned a verdict (two unmerged worktrees, rnd/structure-mapping-experiment and rnd/sme-experiment-v2; the latter's tip is "formalize §5 experiment scaffolding"). So the condemned mechanism keeps serving under an explicit ruling that it should, until a proven replacement exists. That is superseded-in-place, and the schema's separation of liveness from fitness is what lets this card state it without contradiction.

Honest wrinkles:

• The "34 surface organs" count does not reproduce. At this SHA there are 18 resolve_promotable_* functions, all in generate/derivation/. ADR-0252 cites 34. Either the count used a different basis, or consolidation has occurred since ratification. The discrepancy is unresolved and matters, because 34→18 would be evidence the debt is already being paid down — or evidence the ADR's diagnosis was calibrated against something else. Phase 3 should resolve it rather than repeat either number.
• The single load-bearing empirical claim of the governing paradigm ADR is unresolved, and a well-controlled NO-GO is defined as full credit — so the experiment is cheap to finish and expensive to leave open. This is the highest-leverage open item in the assessment.
• Deduction's strength and comprehension's narrowness are easy to conflate. wrong=0 across 18,000 cases is a statement about the decision procedure, not about how much English CORE can read.
• field-wedge-research is research-negative — a mechanism honestly refuted. It is knowledge and must not be tidied into inert.

Open questions:

• Run ADR-0252 §5 to a verdict (→ ruling; highest leverage)
• Reconcile the 34-vs-18 organ count (→ Phase 3)
• Descend the three ⚑ serving-path zones (→ Phase 3, first)
• Is multi-step proof articulation wanted, given it is engine work with a real soundness surface? (→ ruling)