core/docs/research/deduction-serve-arc-phase0-baseline-2026-07-23.md

Deduction-serve arc — Phase 0 baseline (2026-07-23)

Goal of the arc. Ship CORE's first serious end-to-end reasoning workflow:

user asks a simple logic question in `core chat`
  -> comprehend the question into typed premises + query
  -> decide it with the verified propositional/syllogism engine (wrong=0)
  -> articulate the verdict as a grounded, deterministic, provenance-tagged answer
  -> replayable, telemetered, governed by an EARNED serve license

Base SHA: 6a54d27a (main == forgejo/main). Pinned numbers: deduction-serve-phase0-baseline.json (sibling).

This document is the evidence base the rest of the arc rests on. Phase 1 depends on it.

---

The central finding (proven, not asserted)

Every organ this workflow needs already exists and is verified wrong=0 — none of them are connected to each other on the serving path. The arc is wiring + earned licensing, not new cognition machinery.

Demonstrated empirically (probe: comprehend() -> projector -> independent oracle vs ChatRuntime.chat() on the same text):

Input	Organs in isolation	Serving path (ChatRuntime.chat) today
If p then q. p. Therefore q.	to_deductive_logic → ROBDD oracle → entailed	grounding_source='pack', token-gloss surface, refusal_reason=''
All mammals are animals. All whales are mammals. Therefore all whales are animals.	to_syllogism → oracle → valid=True	grounding_source='pack', token-gloss surface

The serving surface for a decidable logic argument is literally "Pack-resident tokens — pack-grounded (en_core_cognition_v1): then (relation.sequence.after; temporal.sequence), therefore (relation.consequence; logic.derivation). No session evidence yet." — the deductive engine is bypassed entirely; the argument is treated as a bag of vocabulary tokens to gloss, not a problem to decide. It is not even an honest refusal (empty refusal_reason).

This is the disconnection the arc closes.

---

Baseline lane numbers (all wrong=0)

Comprehension reader lanes (prose -> comprehend() -> projector -> independent oracle vs gold):

Lane	total	correct	wrong	refused	Band v1?
propositional	12	12	0	0	✅
syllogism	8	7	0	1	✅
set_membership	8	8	0	0	—
total_ordering	8	7	0	1	—
relational_metric	15	15	0	0	—
relational_predicate	18	18	0	0	—
relational_inference	13	13	0	0	—

Deductive engine lane (evals/deductive_logic/report.json, independent truth-table oracle, INV-25):

split	n	correct	wrong	refused
dev	200	200	0	0
holdout_v1	500	500	0	0
external_v1	16	16	0	0
combined	716	716	0	0

all_correct=true, wrong_is_zero=true.

ProofWriter-OWA refusal floor (determine() soundness): 19 total, 9 correct, 0 wrong, 10 refused (acceptable coverage misses), 0 coverage gaps.

Smoke suite (in worktree, cold venv): 180 passed, 0 failed (133s).

Read the numbers honestly: these are small curated gold corpora (8–18 cases per comprehension lane), not a large real corpus. wrong=0 proves the organs are sound on what they read; the counts show the band is narrow. Reader coverage is the wall, not the engine — consistent with the standing project note that coverage is corpus-bound, not architecture-bound.

---

Band v1 — precise input boundary (the key scoping result)

Band v1 = the two endorsed projectors: propositional (to_deductive_logic) + syllogism (to_syllogism). Their natural-language input contract today is narrower than the plan's illustrative example assumed, and the boundary is worth stating exactly, because it decides Phase 1 scope and Phase 4 widening.

The reader (generate/meaning_graph/reader.py) canonicalizes each noun-phrase slot via _chunk (lines 174–190). Its _RESERVED set (lines 109–120: is, are, than, with, not, all, no, some, therefore, the, and, or, …) makes _chunk refuse any multi-token NP that contains a function word — a parse-leak guard that refuses rather than chunk junk.

Propositional band. Reads an argument in P1. P2. Therefore C. declarative form where each proposition/atom is a single content token (p, q, r, or a one-word proposition). It refuses:

• multi-word English propositions like the ground is wet → Refusal('reserved_word_in_np') (contains the, is);
• interrogative propositional Is q true? → Refusal('unreadable_member_query') (the interrogative-propositional path is not wired; only the syllogism reader handles interrogatives).

Syllogism band. Reads All/No/Some X are Y categorical premises with single-word class terms (mammals, whales, animals), in both the declarative Therefore all X are Y form and the interrogative Are all X Y? form. The syllogism reader is the more natural-language-ready of the two.

Consequence for the arc: "natural-English conditionals" (If it rains then the ground is wet…) are out of Band v1 — they need the multi-word-proposition relaxation, which is Phase 4 widening (relaxing _chunk's reserved-word guard without breaking wrong=0), not Phase 1. Band v1 ships the forms that already read: single-token propositional arguments and single-word-term categorical syllogisms.

---

Serving-path seams Phase 1 must touch (verified file:line)

• REPL driver: core/cli.py::cmd_chat constructs ChatRuntime only (not CognitiveTurnPipeline); the turn is ChatRuntime.chat(text). Phase 1 wiring lives in chat/runtime.py, the real core chat path.
• Live comprehension already on the turn path: chat/runtime.py:1342 calls comprehend(text) / comprehend_relational(...), routing to determine(...) or realize_comprehension(...) — but never calls to_deductive_logic / to_syllogism. That is the missing edge.
• Intent: generate/intent.py IntentTag has 12 members; none catches conditional/syllogistic argument shape. A logic argument currently classifies UNKNOWN (or is dispatched to the pack token-gloss composer, as observed). Phase 1 adds IntentTag.DEDUCTION + deterministic classification, field-ratified like every tag.
• Renderer: generate/determine/render.py::render_determination is the sibling pattern; Phase 1 adds a deduction renderer for ENTAILED / REFUTED / UNKNOWN / REFUSED (deterministic templates, no LLM — honesty bar endorsed).
• Surface precedence: docs/specs/runtime_contracts.md "Current selection policy" gains deduction_surface; contract tests updated in the same PR. Trace-hash folds only-when-non-empty so legacy hashes stay byte-identical.
• Fail-closed (INV-34): reader Refusal → typed refusal surface; projector None / oracle REFUSED → honest out_of_decidable_regime; never a fluent-but-ungrounded surface.
• Flag: all Phase 1 behavior behind deduction_serve (default off), byte-identical when off.

What Phase 0 confirms is already handled (no work needed)

• W-012 is fixed: chat/runtime.py:2692-2705 catches InnerLoopExhaustion and materializes a typed refusal_reason. (The zone card's "still lacking except" note is stale.)
• Engine is telemetry-wired but surface-inert: pipeline.py:1203 runs evaluate_entailment_with_trace on precise VERIFICATION triples, folded into operator_invocation only — never the surface. Confirms the engine is trusted evidence, just not a decider on the serving path.

Scope-outs (documented, not on this arc's critical path)

• ADR-0246 §3.7 identity-wave calibration stays OFF — blocked on §11 grounding research evidence (currently NULL; benign false-refusal 1.00; no benign/adversarial separation), not wiring.
• Tier-2 two-reader convergence waits for the field reader (Phase W); today's two registered readers operate on disjoint domains and never co-decide.
• Reader→Hamiltonian compiler stays eval-side (its real-NL reach is 5/500); it is not this workflow's vehicle.

---

Verdict

Phase 0 complete. Baseline pinned, all lanes wrong=0, smoke green, Band v1 boundary defined precisely, the crux disconnection proven with a runnable probe. Proceed to Phase 1 (deduction turn spine behind deduction_serve), scoped to the two forms that already read: single-token propositional arguments and single-word-term categorical syllogisms.