diff --git a/docs/analysis/deductive-logic-phase-2-benchmark-runway-2026-06-04.md b/docs/analysis/deductive-logic-phase-2-benchmark-runway-2026-06-04.md index 0f33f14c..38078d04 100644 --- a/docs/analysis/deductive-logic-phase-2-benchmark-runway-2026-06-04.md +++ b/docs/analysis/deductive-logic-phase-2-benchmark-runway-2026-06-04.md @@ -56,7 +56,16 @@ Do not do these in Phase 2 PR-1: ## 3. Recommended PR stack -### PR-1 — finite-entity grounding contract and tiny mirror fixture +> **Status (2026-06-04): PR-1 and PR-2 SHIPPED** (PR #556), as Phase 2 of the +> universal-structure plan +> (`docs/analysis/universal-structure-and-field-symbol-coherence-gate-2026-06-04.md`). +> `evals/deductive_logic/grounding.py` is the lowering compiler; the committed +> `finite_entity/v1/cases.jsonl` carries oracle-derived gold; the test file gates +> `engine == oracle == gold` (PR-2 parity). PR-3 (published-benchmark mirror) and +> PR-4 (scale + SHA-pin) remain open. The grounding lane is registered under the +> deductive oracle's INV-25 independent-gold coverage. + +### PR-1 — finite-entity grounding contract and tiny mirror fixture — **✅ SHIPPED (#556)** Add the deterministic contract before scaling data. @@ -102,7 +111,7 @@ In v1, the accepted grammar should be intentionally narrow: Refuse any case outside that grammar with a typed reason. -### PR-2 — independent oracle parity for finite-entity cases +### PR-2 — independent oracle parity for finite-entity cases — **✅ SHIPPED (#556)** Before adding benchmark data, assert that the finite-entity lowered formulas score identically under: diff --git a/docs/sessions/SESSION-2026-06-04-deductive-pivot-and-universal-structure.md b/docs/sessions/SESSION-2026-06-04-deductive-pivot-and-universal-structure.md new file mode 100644 index 00000000..27b069ec --- /dev/null +++ b/docs/sessions/SESSION-2026-06-04-deductive-pivot-and-universal-structure.md @@ -0,0 +1,118 @@ +# Session 2026-06-04 — The deductive pivot, the independent-gold invariant, and the universal-structure arc + +**Status:** paused (clean; all work merged to main). **Headline:** Started to +implement a GSM8K Tier-2 / composition plan, discovered through investigation that +the path was inert and the GSM8K composer epistemically unsafe, pivoted (with the +operator) to **checkable deductive logic** as the real verified capability, +**ratified the independent-gold discipline as a structural invariant (INV-25)**, +then designed and began building the **universal comprehension→structure→solve→verify +spine** with the **field↔symbol coherence-gate** synthesis — shipping the +foundation + a 3-domain anti-overfit panel. + +`Five PRs, all gated by independent gold, all green before push: #554 → #557.` +`No serving path, algebra, versor, or recall touched; wrong=0 intact throughout.` + +--- + +## TL;DR + +1. **Investigated the requested 3-PR plan (Tier-2 verifier / target-slot composition / + cross-domain arena) and found it inert.** The Tier-2 spine had **no genuine second + derivation** (solver-vs-verifier shares structure = decoration; the two GSM8K + readers never co-fire). Measured on 500 held-out cases: candidate-graph **5/0**, + `resolve_pooled` **2 correct / 87 wrong**, **0** agreement-flips. The `t2_precision` + "widen past gold" lever was confirmed dormant (nothing consumes it). So the plan + could not safely raise GSM8K, and the composer was *epistemically unsafe as serving*. +2. **A pivot-to-deductive-logic strategy doc arrived; I verified it adversarially + rather than accepting it.** A 6-claim fact-check workflow + 2 adversarial refutations + + my own **20,000-formula differential fuzz (0 mismatches)** confirmed: the ADR-0206 + propositional entailment operator is genuinely **sound + complete**, `wrong=0` + structural, scored against a **genuinely independent** truth-table oracle (holdout + 500/500). The doc's overstatements were caught and corrected (the "8,000-fuzz" / the + "external mirror" mislabel; the stale GSM8K held-out numbers). +3. **Ratified the lesson as Phase 0 (#554, merged): INV-25 — independent gold.** No + capability claim is valid unless its gold shares no code with the system under test. + Three meaningfully-failing checks (proven able to fail). SHA-pinned the deductive lane. +4. **Designed the universal-structure architecture + the field↔symbol synthesis.** The + binding-graph DAG is the universal interlingua; the field comprehends, the structure + bridges, the symbol verifies, and **their agreement is the `wrong=0` gate and the + missing second derivation** — operationalizing "truth is coherent." +5. **`/goal` autonomous build — the foundation + a 3-domain anti-overfit panel:** + Phase 1 (#555, interlingua canonized + **INV-26** neutrality + the ≥2-domain rule); + Phase 2 (#556, the **finite-entity grounding compiler** — first comprehension→structure + compiler); the **dimensional-reasoning lane** (#557 — interlingua's unit algebra as a + reasoner). INV-25 now spans **two** independent oracles. +6. **Two keystone field-as-reasoner findings (why the geometric wedge is deferred to + dedicated research, not rushed):** (a) **logic is combinatorial, not geometric** — a + clean geometric decoder agrees 716/716 but is `O(2^n)`, enumeration-class; (b) + **independence must be in the *reading*, not the *solving*** — two solvers over the + same extracted structure is fake independence. The field must comprehend + *independently*. Rushing it would be decoration. + +--- + +## How we got here (the reasoning trail) + +### 1. The plan as given did not survive contact with the measurements +The session opened with a plan to wire a Tier-2 math verifier (PR 554), then +answer-changing target-slot composition (PR 555), then a cross-domain arena (556). +Reading the tier-2 evidence spine (PR #553) and the reliability gate showed the +verifier would be **inert**: `verify_tier2_agreement` requires ≥2 structurally-distinct +derivations converging on one commitment, and none existed (the ADR-0117 verifier +re-checks the *same* graph + operation DAG). Empirically, **0/6 train + 0/5 holdout** +admitted cases were multi-branch, and the `t2_precision` gate had no consumer. A +deeper probe (candidate-graph vs `resolve_pooled` on 500 held-out) found **0 +agreement-flips** and that `resolve_pooled` commits **87 wrong** — an unsafe serving +reader. The honest conclusion: GSM8K serving is coverage-walled and the composer is +unsound; no safe lift was available. + +### 2. The pivot was earned by verification, not asserted +An external strategy doc argued for pivoting to deductive logic. Rather than accept +its repo claims, a verification workflow fact-checked them against source, two +adversarial agents tried to refute the load-bearing ones, and I ran an independent +20k-case engine-vs-oracle fuzz. All confirmed: deductive logic is the *first real +checkable capability* (sound+complete ROBDD, independent oracle, 500/500 holdout, +wrong=0). The doc's drift was corrected in-repo. + +### 3. The discipline became structure (INV-25), then architecture +The GSM8K breach happened because a single gameable ruler (the 50-case train_sample) +hid unsoundness. INV-25 makes "independent gold" a structural, meaningfully-failing +invariant. From there the architecture followed: **one universal problem-structure +(the binding-graph interlingua) that any reader compiles into, gated by independent +gold + two-derivation agreement** — with the field↔symbol coherence gate as the +synthesis the operator requested (field reasons *only* where it agrees with an +independent decoding; "truth is coherent" made operational). + +### 4. The autonomous build, and where honesty stopped it +Phase 1 canonized the interlingua and added INV-26 (neutrality — the meeting point +must be neutral for agreement there to be real). Phase 2 shipped the first +comprehension→structure compiler (finite-entity grounding). The dimensional lane added +a third structurally-distinct golded domain and proved INV-25 generalizes to a second +oracle. The **anti-overfit ≥2-domain discipline** — the direct antidote to the +train_sample failure — is now live. The field-as-reasoner wedge was investigated +honestly and **deferred to dedicated research** with two hard constraints recorded, +rather than shipped as decoration. + +--- + +## What shipped (all merged) + +| PR | What | +|---|---| +| #554 | INV-25 (independent-gold invariant, 3 meaningfully-failing checks) + SHA-pinned `deductive_logic_v1` lane + drift fixes | +| #555 | Phase 1 — binding-graph interlingua canonized + **INV-26** (neutrality) + diversity-panel / ≥2-domain rule | +| #556 | Phase 2 — finite-entity grounding compiler (`evals/deductive_logic/grounding.py`) + the Phase 1.5 finding | +| #557 | Dimensional-reasoning lane (`evals/dimensional/`) — 3rd panel domain; INV-25 → 2 oracles | + +Plan of record: [`docs/analysis/universal-structure-and-field-symbol-coherence-gate-2026-06-04.md`](../analysis/universal-structure-and-field-symbol-coherence-gate-2026-06-04.md). + +## Open / next (dedicated effort) + +- **Field-reasoner wedge** — independent geometric *comprehension* of a quantitative + domain (not solving). The genuine second derivation that would unblock `t2_precision` + (Phase 3) and the geometric arc. Research-grade; needs `algebra/`+`field/` integration + and a domain/encoding decision. Constraints: logic is out (combinatorial); independence + must live in the reading. +- **Phase 4** — GSM8K on the universal structure, gated by agreement + held-out/sealed + (never train_sample). Depends on the wedge. +- **Phase 5** — systems/software execution-gold arena (a 4th, strongest-verifier domain).