core-labs/core
core-labs/core
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity
core/docs/handoff/ADR-0244-D4-IMPLEMENTATION-PLAN.md
Shay 8ef80daa04 docs(adr-0244-d4): Phase 5 COMPLETE — §2.7/§2.9/§2.10 + ADR-0245 §3 gate 
Plan doc (scope tables + Phase 5 status/resume + progress log + dashboard) and
ADR-0245 status map updated: §2.3 residual done (5a), §2.9 proven (5b), §2.10
audited (5c), §3 acceptance gate all four legs green (5d). Also corrected the
Phase-4 SHA reference (17ec6eee → 918aa843, amend lag). Next: Phase 6 close-out;
all build phases (0-5) complete.
2026-07-17 20:03:02 -07:00

40 KiB
Raw Blame History

ADR-0244 D4 — Wave-Field Identity Manifold: Implementation Plan & Progress Tracker

Owner: Shay (sole worker on this arc) · Authored: 2026-07-17 Arc base: main @ ee38c976 (== forgejo/main) · Worktree: /Users/kaizenpro/Projects/core-adr0244d4 · Branch: feat/adr-0244-d4 Companion ADRs: ADR-0244 (identity consumer) · ADR-0245 (mechanical-sympathy + semantic-rigor foundation — to be committed in Phase 0) Prior audit: adr-0244-cohesion-directive-audit-2026-07-17.md · Q_top proof: evals/adr_0244_qtop_vacuity/

This document is the resume anchor. It is committed to main and updated at every phase boundary. On a cold start: read §0 → §10 status dashboard → the first phase whose status is not ✅, then its Resume notes. Everything needed to continue is here; do not re-derive from the relays or Google-Drive docs (see §3 for why).

0. Purpose & how to resume

This is a multi-phase, multi-commit arc expected to span rate-limit interruptions. Progress is tracked two ways:

  1. This committed doc (durable; survives new sessions) — the source of truth. §10 dashboard + per-phase Status/Resume notes.
  2. The in-session harness task list (convenience; lost across sessions).

Resume procedure:

  • cd /Users/kaizenpro/Projects/core-adr0244d4 (create the worktree from forgejo/main if gone: see §8).
  • Read §10 dashboard. Find the first phase not ✅ DONE. Read its Resume notes.
  • Continue from there. Update Status + append to §9 Progress log at each phase boundary.

1. Scope: the two ADRs and what's already done

ADR-0244 is the identity-layer consumer; ADR-0245 is the mechanical-sympathy + semantic-rigor foundation it sits on. Measured against main @ ee38c976:

ADR-0244 § Mechanism Status entering D4
2.1 Gram-subspace identity projection ❌ unbuilt (identity.py is legacy ADR-0010 heuristic)
2.2 Spectral-leakage fail-closed gate + C_id + IdentityGateRefusal ❌ unbuilt as gate (call site exists at chat/runtime.py:2679, but advisory only)
2.3 Q_top topological charge ✅ resolved → retire from egress (proven vacuous, evals/adr_0244_qtop_vacuity)
2.4 Bracketed-local Fibonacci search + γ_id calibration ◐ search-honesty done (D3); calibration half unbuilt
2.5 Serving-boundary f64→f32 cast contract ✅ done (Phase 4 — serving_cast→ServingState, cognitive_lifecycle.py)
2.6 Rust GP fast-path + parity gate ✅ done (f32 pre-existing + f64 D2, bit-identical)
2.7 Full digests / no default=str / byte-order guard ✅ done (Phase 5a — 4 residual content-id sites widened)
2.8 eigh memoization ✅ done (_cached_eigh, D2)
2.9 Low-discrepancy mode-centroid allocator ✅ done (Phase 5b — recall order-independence proven, not wired; allocator ordinal-reconstructible)
2.10 Fibonacci-word background scheduler ✅ done (Phase 5c — quasi-periodicity claim pinned, not just structure)
ADR-0245 § Mechanism Status entering D4
2.1 PyO3 Rust geometric_product f32 fast-path ✅ done (algebra/backend.py)
2.2 Gated f64→f32 serving boundary ✅ done (Phase 4 — serving_cast, == 0244 §2.5)
2.3 Semantic rigor in content addressing ◐ hot-path done (D1); residual
2.4 _cached_eigh memoization ✅ done
3 Acceptance gate (parity, ≥10× f32 speedup, 0-LAPACK-on-repeat, collision-resistance) ◐ parity ✓, 0-LAPACK partial (test_adr_0244_mechanical_sympathy.py); f32-speedup + collision proofs missing

Net remaining build (entering Phase 5) = §2.7 residual (3 contemplation content-id sites), §2.9 allocator wiring, §2.10 audit-confirm, + the ADR-0245 §3 acceptance gaps (f32 ≥10× speedup benchmark + collision-resistance proof). (§2.1/§2.2 identity gate, §2.4 calibration, §2.5/0245§2.2 cast all landed Phases 1–4.)

2. Ratified decisions (LOCKED — Shay signed off 2026-07-17)

  1. Drop ∧ ΔQ_top = 0 from the §2.2 egress condition. Proven vacuous (Q_top ≡ 0 on every versor; hollow gate). Egress becomes score ≥ threshold (leakage-based). Scope the §2.3 conservation claim to versors only; keep Q_top as a diagnostic, never an admit condition.
  2. §2 governs §4. §4's illustrative per-axis-resonance code is rewritten to the §2.1/2.2 Gram/leakage formulation. Remove the (formerly phantom) "ADR-0245" reference by pointing it at the now-real ADR-0245. Replace the bare assert byte-order guard with a typed guard.
  3. Commit ADR-0245 as a real companion ADR (currently untracked in ~/Downloads). Cross-link 0244 ↔ 0245.
  4. Grade-1 axis lift. Pack direction vectors are dim-3 (verified); lift them to Cl(4,1) 32-vectors by placing the 3 components at the grade-1 e1/e2/e3 blade slots (NOT embed_point, which would make Gram a distance table). Orthonormal axes ⇒ G = I.
  5. Signed overlap — never abs() the per-axis inner product; a large negative overlap is anti-alignment (opposition), tracked distinctly from orthogonality.
  6. Euclidean leakage norm — ‖S_id‖ is the positive-definite coefficient-Euclidean norm sqrt(Σ S_id[k]²), explicitly not the indefinite ⟨S_id, reverse(S_id)⟩₀ (which can be 0 for nonzero leakage under signature (+,+,+,+,−)). Projection geometry uses the indefinite metric Gram; leakage magnitude uses the Euclidean norm.
  7. Activate boundary_ids — hard-boundary evaluation, currently dormant (check() iterates value_axes only).
  8. Frozen identity manifold — axis eigenmodes are fixed at pack load and never mutated within a session. Biography holonomy (H_bio ← H_bio·R) accumulates separately and does not rewrite the identity subspace. This preserves inalienability-by-construction.
  9. Bounded, abstaining C_id — the corrector may make a bounded corrective displacement toward the manifold; if it cannot recover alignment within the bound it abstains (emits IdentityGateRefusal, live params unchanged). It must NOT arbitrarily rewrite reasoning to force a low leakage score ("good-metric / bad-cognition" guard).
  10. Paraphrase-invariance is empirical, not automatic. ADR wording reworded to: "invariant under transformations that preserve the represented trajectory's identity-relevant field geometry; paraphrase robustness is measured across the encoder+propagation pipeline." The eval must measure it, not assert it.

3. Ground-truth facts (code-verified against main @ ee38c976)

Verified by reading the tree, not the relays (which contain at least one hallucination — LanceDB — and a false "identity.py already refactored on the VM" claim; disk shows identity.py unchanged legacy).

  • Fact A — axes are R³ 3-vectors. packs/identity/default_general_v1.json ships direction = dim-3 unit vectors; loader enforces _DIRECTION_LEN. §2.1 assumes 32-vector eigenmodes → requires the grade-1 lift (decision #4).
  • Fact B — two field surfaces, not one. §2.2 gate runs on result.final_state.F (cognition-pipeline field, chat/runtime.py:2679). §2.5 ψ_steady cast lives in core/physics/cognitive_lifecycle.py (ADR-0243 lifecycle), which chat/ + core/cognition/ do not import (A-04 off-serve). The ADR conflates them; they are separate contracts in separate files.
  • Fact C — final_state.F is a VERSOR (operator), not a state vector (field/state.py invariant versor_condition(F) < 1e-6). It is even-grade (grades 0,2,4) with exactly zero grade-1 content (verified). This makes §2.1/§2.2's literal "project ψ onto the grade-1 subspace" vacuous (P_id(F)=0, flags everything). Resolution (ratified 2026-07-17): operator-preservation — measure whether the versor preserves the value subspace via F aᵢ F̃. See ADR-0244 governance annotation item 12 + §4a. This is the D4 core-mechanism correction.
  • Advisory, not a gate. Current identity_score feeds _build_surface_context (hedge/claim-strength, runtime.py:2705) + telemetry. would_violate is never called in runtime. §2.2 converts advisory → fail-closed = a live-serving behavior change (highest risk).
  • psi_traj is never populated. _make_trajectory_from_result (runtime.py:385) builds from result.trajectory or (result.final_state,); no psi_traj attr. The wave-field 32-vector is available as final_state.F (used by versor_condition(result.final_state.F)). Phase 2 must thread it.
  • boundary_ids dormant — stored on the manifold, never evaluated (decision #7).
  • Filename drift — ADR-0245 + relays reference multimodal_lifecycle.py; it does not exist. The real file is cognitive_lifecycle.py.
  • Surrounding substrate already present — wave_manifold.py, chiral_gate.py, trajectory_invariants.py, fibonacci_search.py, atlas_packing.py, fibonacci_word_schedule.py. identity.py is the one unpromoted file.

4. Folded-in review considerations (from the external grounded critique)

All code-verified or sound; woven into the phases below.

  1. Signed overlap (decision #5) → Phase 1/2.
  2. Euclidean leakage norm (decision #6) → Phase 1.
  3. Paraphrase-invariance empirical (decision #10) → Phase 0 wording + Phase 2 eval.
  4. Bounded/abstaining C_id (decision #9) → Phase 2.
  5. boundary_ids activation (decision #7) → Phase 2.
  6. 5-layer inalienability framing (algebraic / runtime / pipeline / operational / semantic) → Phase 0 ADR rigor.
  7. Richer identity verification suite: ablation (legacy-only vs wave-only vs dual — prove incremental detection value), fail-closed malformed-ψ (NaN / wrong-dim / nonfinite / missing-cert / stale-manifold), near-singular-Gram mode-basis, metamorphic, conservation → Phase 2/3.
  8. TurnEvent telemetry: psi_leakage_norm: float = 0.0, wave_mode_active: bool = False → Phase 2.

5. Phase plan

Dependencies: 0 → {1, 4} · 1 → 2 → 3 · 5 after 0 · 6 last. Each phase = its own gated commit (in-worktree smoke → fast lane for code phases → push feat/adr-0244-d4:main, per Shay's direct-to-main authorization for this arc). No auto-merge.

Phase 0 — Governance reconciliation (ADR edits, NO code)

Objective: make the spec self-consistent so nothing downstream is built on a contradiction. Files: docs/adr/ADR-0244-...md, new docs/adr/ADR-0245-cga-unification-mechanical-sympathy-and-semantic-rigor.md, docs/analysis/adr-0244-cohesion-directive-audit-2026-07-17.md. Steps:

  • Commit ADR-0245 (from ~/Downloads/ADR-0245-*.md) as Proposed; cross-link 0244↔0245; add a status map (0245 §2.1/2.4 done, §2.3 hot-path done, §2.2 open).
  • ADR-0244 §2.2: drop ∧ ΔQ_top = 0; scope §2.3 conservation to versors + cite the proof.
  • ADR-0244 §2.1/§2.2: record signed overlap (#5), Euclidean leakage norm (#6), the leakage-ratio score 1 − ‖S_id‖/‖ψ‖, grade-1 lift (#4); rewrite §4 to Gram/leakage; typed guards + isfinite (no bare assert); point the ADR-0245 reference at the real ADR.
  • Reword paraphrase-invariance (#10); add 5-layer inalienability framing (#6); label scripture by translation + math-as-analogy disclaimer; note boundary_ids activation + frozen-manifold decision + Fact A/B + filename drift.
  • Update audit doc: mark Q4 (Q_top) + the §4-vs-§2 blocker RESOLVED; map 0245 status. Acceptance: both ADRs internally consistent; no ΔQ_top egress conjunct; no phantom refs; no bare assert in normative code blocks; provenance-guard test still green (docs-only, status stays Proposed — flips happen in Phase 6). Gate: smoke 176 (docs-only, but run it) + provenance/ADR pins. Status: ✅ DONE — landed ad37d03b Resume notes: ADR-0244 governance annotation expanded to 11 items (§4a added, supersedes §4, verbatim R&D body preserved). ADR-0245 committed Proposed at docs/adr/ADR-0245-cga-unification-mechanical-sympathy-and-semantic-rigor.md with its own governance annotation + status map. Audit doc wrinkles 1–2 + Q5 line updated to resolved. Gate: smoke 176 passed + provenance/ADR pins (governance_p12, topological_quarantine, third_door_cohesion) 30 passed. Both ADR status lines confirmed still Proposed (no flip). Phase 1's lift_axis spec confirmed against the real algebra.cl41.basis_vector helper (grade-1 lives at component indices 1–5; e1/e2/e3 = basis_vector(0..2)) — no further verification needed before Phase 1 starts.

Phase 1 — §2.1 identity manifold primitive (OPERATOR-PRESERVATION, pure, off-path) — TDD

Objective: the metric-exact operator-preservation primitive, no runtime wiring. (§2.1/§2.2's literal "project ψ onto I" is vacuous — the trajectory final_state.F is a versor/operator (grades 0,2,4, zero grade-1); see ADR-0244 governance annotation item 12 + §4a. We measure whether the versor preserves the value subspace via F aᵢ F̃.) Files: new core/physics/identity_manifold.py (keep identity.py as compat shell + dual-mode host, wired in Phase 2); tests/test_adr_0244_identity_manifold.py. Steps (implement §4a verbatim shape):

  • lift_axis(direction3) → ψ_axis(32): grade-1 embedding via algebra.cl41.basis_vector(0..2) (e1/e2/e3, component indices 1/2/3). Value subspace I lives in spatial grade-1 where ⟨·,·⟩₀ = Euclidean ⇒ G positive-definite.
  • gram_matrix(axes) → G (G_ij = scalar_part(gp(ψ_i, reverse(ψ_j)))), symmetric PD; cond(G) > 1e5 → ManifoldConditioningError.
  • subspace_project(x, axes, Ginv) → P_I(x) = Σ ψ_i (G⁻¹)_ij ⟨ψ_j, x⟩₀ (signed coeffs).
  • sandwich(R, x) = gp(gp(R,x), reverse(R)); euclidean_norm(s) = ‖s‖₂.
  • axis_response(R, axes, Ginv) → (leakage[], self_align[]): per axis rot=sandwich(R,aᵢ), leakage_i=euclidean_norm(rot − P_I(rot)) (subspace departure), self_align_i=⟨aᵢ, rot⟩₀ (signed orientation). Both required, non-redundant. Acceptance (falsifiable, all empirically pre-verified): orthonormal default pack ⇒ G=I; subspace_project idempotent (P_I∘P_I = P_I to 1e-12); identity versor ⇒ all leakage≈0, all self_align≈+1; rotor within value plane (e12) ⇒ leakage≈0; rotor tilting a value axis toward e4/e5 ⇒ leakage>0 (e.g. e14 θ=0.5 → 0.28, boost e15 θ=0.5 → 0.29); π-rotor inverting a value axis (e1→−e1) ⇒ leakage≈0 but self_align≈−1; near-parallel synthetic axes ⇒ ManifoldConditioningError. Deterministic (f64). Gate: smoke + fast lane + new tests. Status: ✅ DONE — landed 3c3d2c29 Resume notes: Shipped core/physics/identity_manifold.py (lift_axis, gram_matrix+ManifoldConditioningError, subspace_project, sandwich, euclidean_norm, frozen IdentityManifoldGeometry with .project/.axis_response/.leakage_rms) + tests/test_adr_0244_identity_manifold.py (24 tests). Pure algebra.cl41, f64, deterministic; off-serve (serve-quarantine test confirms). Gate: 24 new + smoke 176 + fast lane 11863 passed/109 skipped. NOT yet wired into identity.py or the barrel — that's Phase 2.

Phase 2 — §2.2 fail-closed gate + C_id + boundary_ids + telemetry + eval — TDD

Objective: convert advisory → fail-closed egress gate on final_state.F; wire in-path behind a flag. Files: core/physics/identity.py (dual-mode _axis_projection/check, IdentityGateRefusal, C_id), chat/runtime.py (thread final_state.F as psi_traj; act on verdict behind config flag; TurnEvent fields), evals (evals/adversarial_identity, teaching_injection_resistance, identity_divergence), tests. Steps:

  • Dual-mode: psi_traj present → wave-field projection (Phase 1 primitive); absent → legacy ADR-0010 fallback (back-compat). psi_traj present but malformed (NaN/nonfinite/wrong-shape) → typed error, never silent legacy.
  • Score = 1 − leakage_norm/‖ψ‖; flagged if < manifold.alignment_threshold; deviation_axes from signed per-axis contributions (opposition ⇒ deviation).
  • boundary_ids hard evaluation (design the violation predicate; likely ties to safety-pack boundaries — scope in-phase).
  • C_id: bounded corrective displacement; abstain (IdentityGateRefusal, params unchanged) if unrecoverable.
  • Wire runtime.py:2679 as fail-closed behind a config flag defaulting to current advisory behavior until Phase 3 calibrates. Add TurnEvent.psi_leakage_norm, wave_mode_active. Acceptance: all three identity eval suites still pass; the gate demonstrably catches the injections; ablation shows wave path adds detection value over legacy; fail-closed tests (NaN/dim/nonfinite/missing-cert/stale-manifold) never fall through permissively; C_id bounded + abstains on unrecoverable; byte-identity of non-identity turns preserved (flag default off). Gate: smoke + fast lane + eval suites + new tests. Decomposition (sub-commits for clean checkpoints): 2a pure identity.py capability (dual-mode wave path + IdentityGateRefusal + boundary predicate + C_id) — no runtime change; 2b runtime wiring behind identity_wave_gate flag + telemetry; 2c eval suite (ablation / fail-closed / injection-catching). Key facts locked from recon: final_state.F is the wave field (versor); boundary_ids = union of identity+safety(no_identity_override,…)+ethics IDs, so boundary activation is an intersection with the turn's SafetyVerdict.violated_boundaries ∪ EthicsVerdict.violated_commitments (available only AFTER the verdicts, ~runtime.py:2740, not at the 2679 identity call); compute_trace_hash excludes telemetry fields (replay-safe); the JSONL serializer is explicit getattr per field, so wave telemetry lives on IdentityScore and is emitted only when wave_mode_active (byte-identical when off); C_id v1 = admit-or-abstain (no silent correction). Deviation from the reviewer's TurnEvent.psi_leakage_norm suggestion: telemetry lives on IdentityScore instead (serializer already reads it; cleaner + byte-identity-preserving). Status: ✅ DONE — 2a 1c7ea26e, 2b c7e2b3b6, 2c c0ff4720. Resume notes: Phase 2 landed in three gated sub-commits. 2a (1c7ea26e): pure identity.py capability (dual-mode check(..., *, wave_field=, violated_boundary_ids=), _validate_wave_field fail-closed, _wave_field_score, extended IdentityScore, IdentityGateRefusal, admit-or-abstain conjugate_correct, would_violate extended; 16 tests). 2b (c7e2b3b6): runtime wiring behind identity_wave_gate flag (core/config.py) — wave_field=final_state.F at the identity check + post-verdict boundary supplement + geometric refusal fold in chat/runtime.py; wave telemetry in chat/telemetry.py (only when wave_mode_active). Integration fix: per-axis normalization in identity_manifold.py (boost/e5 versors break Euclidean-norm preservation) → leakage ∈ [0,1], self-align ∈ [−1,1]. Runtime test seeds state to reach main-path turns. 2c (c0ff4720): evals/adr_0244_identity_gate/ detection-value ablation — the wave gate separates a geometric-attack panel (tilts/boosts/inversions) from an aligned panel (all attacks flagged, all aligned admitted, margin > 0) and adds detection value 6-vs-0 over the geometry-blind legacy path; 6 tests + CLI. Gate: smoke 176 + fast lane 11883→11889/109. Carried to Phase 3: value axes are placeholders (e1/e2/e3); the gate discriminates the designed geometric attack signal, but whether real paraphrased injections induce such geometry through the live encoder is empirical (governance annotation item 6) — Phase 3 γ_id calibration over reference traces is where that is measured, and the flag stays OFF until it is evidenced.

Phase 3 — §2.4 γ_id calibration

Objective: replace the hardcoded threshold with a certifiable, calibrated bound. Files: calibration eval under evals/, uses core/physics/fibonacci_search.py (bracketed-local, already built); tests. Steps: define a BoundedUnimodalObjective over reference traces (ID traces pass, adversarial fail); run the Fibonacci search; emit an audit-logged tuning certificate; pin the calibrated threshold; flip the Phase-2 flag on once evidenced. Acceptance: calibrated threshold separates ID vs adversarial reference sets; certificate reproducible + deterministic; flag-on run keeps eval suites green. Gate: smoke + fast lane + calibration tests. Status: ✅ DONE — landed f3702ad4. The flag flip is NOT authorized (honest empirical finding); flag stays OFF. Resume notes: Shipped evals/adr_0244_gamma_calibration/ (run_gamma_calibration, calibrate_leakage_bound, leakage_separation_objective, collect_live_benign_leakages, CLI) + tests/test_adr_0244_gamma_calibration.py (9 tests, 1 slow drift-guard). Pinned identity._WAVE_LEAKAGE_BOUND = 0.2126624458513829 (cert 0079b5f201fbf616…, objective gamma_id_leakage v1, budget 24, interval [0,1], sharpness 10) and decoupled the wave path from alignment_threshold (legacy path + hedge bands keep it; wave path now flags on leakage_rms > _WAVE_LEAKAGE_BOUND). The self-alignment floor stays 0.0 (geometric invariant, not calibrated).

  • Two honest results, kept separate: (1) the bracketed-local Fibonacci search certifies γ* separating the geometric attack signal — every in-subspace rotor admitted, every axis→e4/e5 tilt/boost flagged (inversions excluded from the leakage set — they're ~0-leakage, caught by the orientation floor); machinery validated + reproducible + deterministic. (2) The same bound does NOT separate real live traffic. Measured on the arc engine (ChatRuntime, wave flag on, 13 main-path benign turns): benign final_state.F leakage spans 0.14–0.81 (mean 0.55), self-alignment swings to −0.52; 12/13 benign turns would be false-refused at γ*, benign overlaps the attack range, and the best achievable balanced error over all thresholds is 0.346. No leakage bound gates live traffic without mass-refusing benign turns.
  • Root cause (architectural gap, not an impl defect): the shipped pack value axes (truthfulness=e1, coherence=e2, reverence=e3) are nominal basis vectors, not the dynamically-preserved eigenmodes §2.1 presumes. The current field evolution gives the identity subspace no dynamical anchoring, so an ordinary cognition versor rotates it freely. Closing this = making identity dynamically load-bearing = the ADR-0246 induced-identity-action programme (preflight brief already merged, docs/briefs/ADR-0246-*). The wave gate is therefore validated, correctly flag-gated-off scaffolding, not a live gate. A slow drift-guard test re-measures the live distribution and fails loudly if the engine ever starts preserving identity (→ re-calibrate + reconsider the flip). This resolves the last open question from §7 (boundary_ids predicate was Phase 2; the live-separation question was the real Phase-3 crux). The eval's CLI exit-0 means "certificate valid", explicitly not "flip the flag".

Phase 4 — §2.5 / ADR-0245 §2.2 serving-boundary cast (lifecycle-internal, independent)

Objective: governed explicit f64→f32 cast at the certified lifecycle egress. Files: core/physics/cognitive_lifecycle.py (cast psi_steady→f32 at the certified boundary), tests. Steps: add the typed cast contract at the egress hand-off; precision-sufficiency + parity test; keep f64 inside relaxation/eigendecomp. Acceptance: f32 cast only after certification; parity within documented f32 tolerance; f64 preserved upstream; determinism intact. Satisfies both 0244 §2.5 and 0245 §2.2. Gate: smoke + fast lane + cast tests. Status: ✅ DONE — landed 17ec6eee. Resume notes: Shipped serving_cast(psi_steady, certificate, verdict, *, tol=1e-6) → ServingState + ServingCastError in core/physics/cognitive_lifecycle.py (+ __all__), pinned by tests/test_adr_0244_serving_cast.py (10 tests). The single explicit f64→f32 down-cast at the certified egress hand-off: fail-closed — casts only a state that validates as a finite 32-vector, matches its certificate psi_digest, and was verdict.admitted; then precision-checks the f32 result (cast_error = max|f64−f32|, unit-norm) and fails closed on an f32 cliff (f32_precision_insufficient). f64 stays source of truth — psi_steady + the digest chain are untouched; ServingState carries provenance back (source_psi_digest, certificate_id, measured cast_error) for audit. Measured on a real certified outcome: cast_error ≈ 1.2e-8, unit_norm_f32 ≈ 0.99999998 (well within tol). Off-serve (A-04 guard test: chat.runtime never imports it); f64 kept inside relaxation/eigendecomp. Independent of 1–3.

Phase 5 — §2.7 residual + §2.9 adoption + §2.10 audit + ADR-0245 §3 gate

Objective: finish semantic-rigor residuals, wire the allocator, close the 0245 acceptance gate. Files: core/contemplation/schema.py, plan_preflight.py, miners/articulation_quality.py (full 256-bit digests + typed coercion); core/physics/holographic_vault.py (route standing-wave-mode centroid via atlas_packing.golden_angle_pack, or prove existing registration insertion-order-independent); new benchmark + collision tests. Steps:

  • §2.7: full-digest + typed-coercion the three contemplation content-id sites.
  • §2.9: wire/verify the allocator into standing-wave-mode registration.
  • §2.10: audit-confirm fibonacci_word_schedule.py satisfies the spec (no build).
  • ADR-0245 §3: add f32 ≥10× speedup benchmark + _content_id collision-resistance proof (parity + 0-LAPACK already covered). Acceptance: no truncated machine merge keys in contemplation content-ids; allocator insertion-order-independent + wired; 0245 §3 four assertions all green. Gate: smoke + fast lane + new tests. Status: ✅ DONE — landed 5a 5c69b741 · 5b 4a550c1d · 5c eacea087 · 5d b31d6fde (+ this docs commit). Resume notes: Four gated sub-commits.
    • 5a §2.7 (5c69b741): widened four residual content-id sites to full 256-bit digests — schema._sha256_16→_content_digest (finding_id/run_id), plan_preflight._plan_substrate_hash, miners/articulation_quality._substrate_hash_for_observations (also dropped default=str — as_dict is JSON-native so a non-serializable field now fails closed), and (a 4th site found in-flight, folded in) holographic_vault._default_mode_id (full digest + explicit LE byte-order, was bare .tobytes()). Content-addresses only → no persisted-id migration; 334 contemplation/vault/miner tests green (they assert determinism + content, not id length).
    • 5b §2.9 (4a550c1d): proven, not wired. tests/test_adr_0244_mode_order_independence.py (3 tests) pins that WaveManifold recall (compute_spectral_leakage → _metric_project Gram lstsq onto span(modes)) is insertion-order-independent to ~1e-15 (machine eps; nonzero-residual probe too), and atlas_packing.golden_angle_pack is ordinal-reconstructible. Wiring the allocator into the content-seal path would be a category mismatch (that path seals content ψ, not packing positions).
    • 5c §2.10 (eacea087): audit-confirm. tests/test_adr_0244_scheduler_quasiperiodicity.py (32 cases) pins the load-bearing "reduce harmonic phase-locking" claim (golden-ratio letter density A/B→φ, no BB/no AAA clustering, aperiodicity), not just the recurrence structure the D-era test already covered. No production change.
    • 5d §3 (b31d6fde): ADR-0245 §3 gate closed. tests/test_adr_0245_acceptance_gate.py (6 tests) — Rust-guarded f32 ≥10× speedup (measured 467×: rust 2.86µs vs python 1340µs/op; skips-with-reason where core_rs unbuilt) with a parity sanity check, + _content_id collision-resistance (full 64-hex, type/structure-faithful, fail-closed on non-serializable, _psi_digest sub-epsilon sensitive), + a coverage manifest mapping all 4 §3 legs. Note: core_rs was built in this worktree venv (maturin develop --release -m core-rs/Cargo.toml) to run the speedup + the ~126 Rust-parity tests live — all green; the build is a local dev artifact (not committed; backend still falls back to Python by default).

Phase 6 — Close-out

Objective: acceptance packets + ratified status flips + memory + cleanup. Files: docs/audit/adr-0244-acceptance-packet-*.md, docs/audit/adr-0245-acceptance-packet-*.md; ADR status flips; memory. Steps: D10-pattern acceptance packets for both ADRs; user-ratified Proposed→Accepted flips (provenance guard: inline **Accepted** — ratified by Joshua Shay <date> + packet ## 8. RULING RECORD); update memory; delete worktree + branch (merge-then-cleanup). Acceptance: both packets complete; provenance-guard green; status flips only on explicit user ratification; worktree/branch swept. Gate: smoke + provenance/ADR pins. Status: ⬜ NOT STARTED Resume notes: —

6. Out-of-scope backlog (do NOT build in D4)

Forward-looking mechanical-sympathy items from the critique — a separate optimization backlog, not ADR-0244/0245:

  • SIMD 64-byte alignment of the Cl(4,1) array.
  • Relaxation-propagator precompute cache.
  • Columnar TurnEvent audit storage.
  • biography.py holonomy versioned storage.
  • Parallel Fibonacci branch exploration.

(Gap-3 psi_digest byte-order is already handled by D1's LE guard.)

7. Open questions

  • ✅ Q_top vacuity — RESOLVED (proven, evals/adr_0244_qtop_vacuity).
  • ✅ §4-vs-§2 reconciliation — RESOLVED (decision #2; §2 governs).
  • ✅ Identity-continuity (mutate vs frozen) — RESOLVED (decision #8; frozen manifold).
  • ⚠️ boundary_ids violation predicate — how a trajectory "violates" a stored boundary id is undefined; design in Phase 2 (likely ties to safety-pack boundaries).

8. Workflow & env notes

  • Worktree: /Users/kaizenpro/Projects/core-adr0244d4, branch feat/adr-0244-d4, based on ee38c976 (== forgejo/main). Recreate if gone: git worktree add -b feat/adr-0244-d4 /Users/kaizenpro/Projects/core-adr0244d4 forgejo/main (fetch forgejo first).
  • Push target: forgejo remote; land phases via git push forgejo feat/adr-0244-d4:main (Shay authorized direct-to-main for this sole-worker arc), then ff local main. Forgejo→GitHub origin auto-mirrors.
  • Local-first CI gates (per phase, before push): in-worktree smoke uv run core test --suite smoke -q (176, doctrinal merge gate) → for code phases, fast lane uv run python -m pytest -n auto -m "not quarantine and not slow" (~9–10 min). [Verification]: line in every commit body.
  • Worktree venv: real (not symlinked). First code phase: uv sync in the worktree. core_rs build (if algebra touched): uv run --project /Users/kaizenpro/Projects/core-adr0244d4 maturin develop --release (plain VIRTUAL_ENV+PYO3_PYTHON trips maturin cross-compile detection).
  • A-04 serve quarantine: new capability code goes in evals/; in-place core refactors keep the off-serve quarantine (no imports into chat/runtime.py from off-serve modules).
  • No auto-merge / no self-Accept. Status flips only on explicit Shay ratification.

9. Progress log (append-only)

  • 2026-07-17 — Arc kicked off. Plan doc authored + committed to main. Worktree core-adr0244d4 created off ee38c976. Both sign-offs recorded (§2). Next: Phase 0.
  • 2026-07-17 — Phase 0 landed ad37d03b. ADR-0244 reconciled (11-item governance annotation + new §4a superseding §4); ADR-0245 committed as a real companion ADR (Proposed) with its own status map; audit doc updated. Gate green (smoke 176 + provenance/ADR pins 30). Pushed to forgejo/main, local main fast-forwarded, worktree in sync. Next: Phase 1.
  • 2026-07-17 — Phase 1 core-mechanism correction (ratified), amendment 459280e3. Discovered before implementing: final_state.F is a versor/operator (grade-1 energy exactly 0), so §2.1/§2.2's literal grade-1 projection is vacuous (flags everything). Ratified switch to operator-preservation (sandwich F aᵢ F̃): subspace-rejection leakage (catches e4/e5 tilt) + signed self-alignment (catches in-subspace inversion). Both empirically verified necessary + discriminating (identity→0 leak/+1 align; e14/e15 tilt→0.28/0.29 leak; π-invert→0 leak/−1 align). ADR-0244 governance annotation item 12 + §4a rewritten; plan §3 Fact C + §5 Phase 1 updated.
  • 2026-07-17 — Phase 1 landed 3c3d2c29. core/physics/identity_manifold.py (operator-preservation primitive) + tests/test_adr_0244_identity_manifold.py (24 tests). Pure/f64/deterministic, off-serve (quarantine test green). Gate: 24 new + smoke 176 + fast lane 11863 passed, 109 skipped (9:01). Not yet wired into serve — Phase 2. Next: Phase 2.
  • 2026-07-17 — Phase 2a landed 1c7ea26e. Pure identity.py capability: dual-mode wave gate + IdentityGateRefusal + boundary predicate + admit-or-abstain C_id + extended IdentityScore + 16 tests. No runtime behavior change (legacy path byte-identical). Gate: 16 new + smoke 176 + fast lane 11879 passed, 109 skipped (8:39). Next: 2b (runtime wiring behind identity_wave_gate flag).
  • 2026-07-17 — Phase 2b landed c7e2b3b6. Runtime wiring: identity_wave_gate flag (core/config.py), two flag-gated touches in chat/runtime.py (wave_field=final_state.F at the identity check; post-verdict boundary supplement + geometric refusal fold), wave telemetry in chat/telemetry.py (emitted only when wave_mode_active). Integration bug caught + fixed: the live versor carries boost (e5) components → F aᵢ F̃ is NOT Euclidean-norm-preserving → un-normalized leakage/self-align ran to 5.16 / −4.75. Fixed with per-axis normalization in identity_manifold.py (leakage fraction ∈ [0,1], signed cosine ∈ [−1,1]; Phase-1 unit-rotor results unchanged; boost test added; §4a updated). Gate: 74 targeted + smoke 176 + fast lane 11883 passed, 109 skipped (8:26) — flag-off byte-identity confirmed. Open reality for Phase 3: value axes are placeholders (e1/e2/e3), so whether the geometric gate DISCRIMINATES adversarial vs benign is an empirical question — 2c measures it. Next: 2c (ablation eval).
  • 2026-07-17 — Phase 2c landed c0ff4720 → Phase 2 COMPLETE. evals/adr_0244_identity_gate/ (detection-value ablation) + tests/test_adr_0244_identity_gate_eval.py (6 tests) + CLI. Measured: the wave gate separates a geometric-attack panel (2 inversions caught by orientation, 4 tilts/boosts caught by subspace leakage) from an aligned in-subspace panel — all attacks flagged, all aligned admitted, min_attack_signal 0.35 > max_aligned_leakage 0.0; wave adds detection value 6-vs-0 over the geometry-blind legacy path. Honest caveat baked in: separation is on the designed geometric signal; real-encoder separation is empirical (Phase 3). Additive-only (evals/ + test), off-serving (quarantine green). Gate: smoke 176 + fast lane. Next: Phase 3 (γ_id calibration).
  • 2026-07-17 — Phase 3 landed f3702ad4 → calibration built + certified; live flag flip BLOCKED by empirical non-separation. evals/adr_0244_gamma_calibration/ + tests/test_adr_0244_gamma_calibration.py (9 tests). The bracketed-local Fibonacci search certifies γ* = 0.2126624458513829 (cert 0079b5f201fbf616…) separating the geometric attack signal; pinned as identity._WAVE_LEAKAGE_BOUND and decoupled the wave path from alignment_threshold. Then measured the crux the whole arc deferred: real benign live turns do NOT preserve span(e1,e2,e3) — leakage 0.14–0.81 (mean 0.55), self-align to −0.52, 12/13 benign false-refused at γ*, best-achievable balanced error 0.346. So the calibration certifies flag_flip_authorized=False; identity_wave_gate stays OFF. Root cause: the shipped value axes are nominal basis vectors, not dynamically-preserved eigenmodes — no dynamical anchoring in the current field evolution; the fix is the ADR-0246 induced-action programme (brief already merged). The wave gate is validated, correctly-off scaffolding. A slow drift-guard test canaries any future change that starts preserving identity. Gate: smoke 176 + fast lane + 59 targeted (incl. wave/ablation/runtime unchanged). Next: Phase 4 (§2.5 cast) — independent of 1–3.
  • 2026-07-17 — Phase 4 landed 918aa843 → §2.5 / 0245 §2.2 serving cast COMPLETE. serving_cast(psi_steady, certificate, verdict) → ServingState + ServingCastError in core/physics/cognitive_lifecycle.py; tests/test_adr_0244_serving_cast.py (10 tests). The single explicit, fail-closed f64→f32 down-cast at the certified egress: casts only a validated / digest-matched / admitted state; precision-checks the f32 result and fails closed on a cliff (f32_precision_insufficient); keeps f64 as source of truth (digest chain untouched) and carries audit provenance. Measured cast_error ≈ 1.2e-8. Off-serve (A-04 guard green). Gate: smoke 176 + fast lane + 10 targeted. Next: Phase 5 (§2.7 residual + §2.9 allocator + §2.10 audit + 0245 §3 gate).
  • 2026-07-17 — Phase 5 landed (5a–5d) → all D4 build phases COMPLETE. Four gated sub-commits closing the semantic-rigor + mechanical-sympathy tail. 5a §2.7 (5c69b741): four residual content-ids widened to full 256-bit digests (schema/plan_preflight/articulation_quality — dropped default=str — + holographic_vault _default_mode_id with explicit LE byte-order); 334 contemplation/vault tests green. 5b §2.9 (4a550c1d): recall order-independence proven (Gram lstsq onto span is permutation-invariant to ~1e-15; allocator ordinal-reconstructible) — wiring would be a category mismatch. 5c §2.10 (eacea087): the "reduce harmonic phase-locking" claim pinned (golden-ratio density, no BB/AAA, aperiodicity), not just structure. 5d §3 (b31d6fde): ADR-0245 §3 gate closed — Rust-guarded f32 ≥10× speedup (measured 467×) + _content_id collision-resistance + coverage manifest. Built core_rs locally to run the speedup + ~126 Rust-parity tests (all green; not committed). Gate: smoke 176 + fast lane + 43 targeted. Next: Phase 6 (close-out) — the only remaining work; all §-mechanisms built.

10. Status dashboard

Phase Objective Status Landed at
0 Governance reconciliation (ADR-0244 edits + commit ADR-0245) ✅ DONE ad37d03b (+ 459280e3 §4a amend)
1 §2.1 identity manifold primitive (operator-preservation) ✅ DONE 3c3d2c29
2 §2.2 fail-closed gate + boundary_ids + C_id + telemetry + eval ✅ DONE 1c7ea26e+c7e2b3b6+c0ff4720
3 §2.4 γ_id calibration (certified; live flip BLOCKED — flag stays OFF) ✅ DONE f3702ad4
4 §2.5 / 0245 §2.2 serving-boundary cast ✅ DONE 918aa843
5 §2.7 residual + §2.9 wiring + §2.10 audit + 0245 §3 gate ✅ DONE 5c69b741+4a550c1d+eacea087+b31d6fde
6 Close-out (2 acceptance packets + 2 ratified flips) ⬜ NOT STARTED —

▶ NEXT: Phase 6 — Close-out. Two D10-pattern acceptance packets (docs/audit/adr-0244-acceptance-packet-*.md, docs/audit/adr-0245-acceptance-packet-*.md) + user-ratified Proposed→Accepted status flips for both ADRs (provenance guard: inline **Accepted** — ratified by Joshua Shay <date> + packet ## 8. RULING RECORD; the anti-self-Accept test forbids me flipping status without explicit ratification) + memory + worktree/branch cleanup. All build phases (0–5) are complete; ADR-0244 §2.1–2.10 and ADR-0245 §2.1–2.4 + §3 are built/proven/audited. The one standing caveat carried into Phase 6 and the packets: identity_wave_gate stays OFF (Phase 3 live-non-separation finding) — the identity gate is validated scaffolding, not a live gate, pending ADR-0246.

Phase 3 headline for resumers: the γ_id calibration machinery is built + certified, but the empirical finding is that the live serving flag cannot honestly be flipped on — real benign traffic doesn't preserve the nominal identity subspace (best balanced error 0.346), so the gate would mass-refuse. identity_wave_gate stays OFF; the live gate awaits the ADR-0246 induced-action work. This is a true architectural-gap discovery, faithfully honoring the plan's conditional ("flip once evidenced"), not a deviation.