core/docs/audit/adr-0243-acceptance-packet-2026-07-17.md

60 lines
6.6 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

# ADR-0243 — D10 Acceptance Packet for Joshua
**Arc:** ADR-0243 Wave-Field Cognitive Lifecycle (Phases 05), built on ADR-0241 (substrate) + ADR-0242 (operators).
**Evidence base:** `feat/adr-0243-phase4-benchmarks @ 2bb57a86` (Phases 04 merged to `main` through #56/#58/#59/#60 + this branch's Phase 4). Becomes `main@<merge-sha>` when Phase 4 merges.
**Status of ADR-0243:** **Accepted**. This packet was submitted *for your ruling*; the `Proposed → Accepted` flip is a **separate one-line governance commit** carrying your ratification provenance inline (per the anti-self-Accept guard, `tests/test_adr_0241_governance_p12.py::test_adrs_accepted_with_recorded_ruling_provenance`). The status flip has been made — §8 is RATIFIED.
---
## 1. Evidence summary (what is proven, and how)
| ADR § | Claim | Built as | Evidence | Verdict |
|-------|-------|----------|----------|---------|
| §2.1 | Ingress = wave ingestion, normalized at one owned boundary | `ingest_context` (superpose via `sensorium_wave_feed`; degenerate superposition refused, never zero-filled) | `test_adr_0243_cognitive_lifecycle` ingress pins; Lane B corridor | **Supported** |
| §2.2 | Reasoning = Hamiltonian well relaxation; "guaranteed, zero intermediate fabrication" | `compile_quadratic_well` / `compile_propositional` + `relax_to_ground` (imaginary-time, D-1); `RelaxationCertificate` (eigen-residual + spectral gap + `psi_digest`) | Phase 4 **fidelity ≥ 0.999**; **insertion cost** all certified + bounded; ground-state property test vs. direct eigensolve | **Supported, honestly restated (D-2)** |
| §2.2 | "The field engine is the reasoner" on checkable logic | `propositional_entails` (ground-energy UNSAT reading) | Phase 4 **decisive falsifier: field vs. ROBDD gold `wrong == 0`** over 1026 decisions (1008 ID + 18 refusal-parity); OOD (>5 atoms) field-refuses / gold-decides | **Supported** on this domain |
| §2.3 | Egress = thermodynamic wave readback, coherence-gated | `egress_gate` (measure_unitary_residual + coherence residual + `EnergyClass` routing); `certificate_state_mismatch` binds cert↔state | egress must-reject pins; E3/E4 → readback-eligible; Lane B corridor readback | **Supported** |
| §2.4 | Speculative contemplation = non-resonant curiosity, proposal-only | surprise→discovery wiring (Lane A `contemplate_surprise_history`); crystallization proposal (I-03) | Lane A discovery pins; crystallization-proposal pin (D-5) | **Supported (proposal-only)** |
| §2.5 | Lifelong learning = biography holonomy update, human-gated | `biography_wiring.integrate_validated_biography` (Lane C); ADR-0240 PASS → integrate seam; I-01 closure at boundary | Lane C report-level PASS gate + I-01 seam + structural no-vault/no-evals pin | **Supported (ratification-gated)** |
| Falsifiability | Metrics separate ID from OOD; f32 truncation is real | Phase 4 `benchmark.py` | **surprise separation 0.834** (> 0.05); **f32 drift** f64 ~8e-14 vs f32 ~4.7e-5 (ratio ~5.9e8) | **Supported** |
The load-bearing result: on the propositional slice — the one domain with an independent, canonical, complete oracle (ADR-0201 ROBDD) — the Cl(4,1) field relaxation decoder agrees with the gold on **every** satisfiable-premise problem (`wrong == 0`), matches its refusal on inconsistent premises, and honestly refuses out-of-regime. "The field engine is NOT the reasoner today" stands unless this evidence says otherwise; it says otherwise, on this domain.
---
## 2. Deviations submitted for ruling (implementation vs. ADR prose)
- **D-1 imaginary-time relaxer** replaces the sketch's `exp(H·I·t)` unitary loop. `exp(H·dt)` via `np.linalg.eigh` on symmetric H (pure numpy f64, no scipy). Imaginary-time (dissipative) descent is the correct operator for relaxation-to-ground; real-time unitary evolution would not settle.
- **D-2 honest guarantee restatement.** §2.2's "mathematically guaranteed … zero room for intermediate fabrication" is **conditional**, holding iff (a) the Hamiltonian compiler encodes constraints faithfully, (b) a positive spectral gap separates the ground state, (c) iteration has **certified** convergence (not assumed). The certificate records exactly these; unresolved cases fail closed (`spectral_gap_below_tolerance`, `excited_eigenspace`, `max_steps_exhausted`), never mis-certified.
- **D-3 I-05 reconciliation.** Relaxation is *deliberately non-unitary* inside the off-serving solver — that is what dissipation is. I-05 (amplitude conservation) governs **live serve-path** transitions; the relaxer's terminal normalize is an **owned construction boundary** (no hot-path drift repair; boundary breaches fail closed).
- **D-4 R-02 premise retired at v1 scale.** `expm` of a 32×32 symmetric matrix in numpy is trivial; no Rust port needed or permitted for v1. *(The ADR-0244 cohesion arc separately revisits a Rust f64 GP fast-path for the algebra layer — a different surface, gated behind `CORE_BACKEND=rust` with bit-identical parity.)*
- **D-5 E0/E1 crystallization is proposal-only.** No vault writes from the lifecycle; cold states emit a typed crystallization **proposal artifact** through the one-mutation-path (I-03), never a direct `COHERENT` write.
---
## 3. Deferred items (explicitly NOT completion blockers)
- Rust §5 SIMD / MLX acceleration of the algebra layer → ADR-0244 cohesion arc (D2), opt-in `CORE_BACKEND=rust`.
- CRDT multi-writer merge semantics → `docs/analysis/crdt-vs-bitexact-determinism-decision`.
- Serve-path wiring of the lifecycle → remains **A-04 off-serving** (Tier-2); no serve consumer at v1.
- Wave-field IdentityManifold / Gram projection / Q_top egress gate → ADR-0244 (§2.12.3), **not** part of this arc; Q_top pending an empirical non-vacuity demonstration (central `I₅` in odd Cl(4,1) makes the naive charge vacuous on versor states — the #19 pseudoscalar failure mode).
---
## 4. Gate results (this arc)
- Phase 2 flagship (#56): smoke 176; fast lane ~11.7k; lane/pins/cohesion/transitive green.
- Phase 3 Lanes C/B/A (#58/#59/#60): each smoke 176 + own tests + fast lane green.
- Phase 4 (this branch, `2bb57a86`): smoke **176**; fast lane **11808 passed, 108 skipped**; Phase 4 tests **17**; benchmark `overall_passed: true` (deterministic); decisive falsifier `wrong == 0`.
---
## 5. Requested action
**Rule on the ADR-0243 arc.** If you rule "Ratify", the follow-up governance commit flips `**Status**: Proposed → **Accepted** — ratified by Joshua Shay <date> (… this packet …)` and records your ruling in §8 below, exactly as the ADR-0241/0242 packet did. The four honest restatements (D-1…D-5) are submitted for acceptance as recommended.
---
## 8. RULING RECORD
**RATIFIED — Joshua Shay**. Status flipped to Accepted on 2026-07-17 with provenance inline.