core/docs/research/intelligence-loop-homestretch-plan-2026-07-18.md

Intelligence-Loop Homestretch Plan

Status: Proposed plan — awaiting Shay's go before any implementation Date: 2026-07-18 Base: main @ 54659b76 (Rings 2+3 merged; ADR-0243/0244/0245 Accepted; ADR-0246/0247/0248 Proposed, flags off) Foundation: docs/research/spark-audit-adjudication-2026-07-18.md — every work item below survived claim-by-claim verification against the live tree; stale audit claims were removed there and must not resurface as work items. Related: ADR-0240, ADR-0243…0248, docs/research/core_cohesion_master_plan.md

---

1. Objective (done-when)

Close the comprehend → reason → articulate → contemplate → learn loop and make its value measurable:

1. All green CI (smoke + fast lanes) at every phase boundary.
2. Meaningful, generalized lift in problem solving, without overfitting — decided by a dedicated instrument (Phase 4), not by narrative. An honest NULL is a legitimate, recordable outcome.
3. Articulate comprehension of inputs and articulate generation of outputs as one cumulative loop — the corridor's field→token readback wired and round-trip-verified (Phase 1).
4. generate/ ("core_logos" role) upgrade explicitly deferred until the above land (Phase 6 scope note).

2. Governance constraints binding all phases

• A-04 serve-path containment: wave-field / Fibonacci operators stay out of chat/runtime.py; corridor work lives in core/physics/ + evals/ quarantine. Serving consumes only ratified frozen artifacts.
• I-03 / proposal-only learning: no autonomous mutation of the active manifold; biography integration only via the PASS-gated wiring layer, harness-driven.
• No self-Accept: ADR-0246/0247/0248 stay Proposed; phases produce acceptance-packet evidence only. Flag flips and Accepts are Shay's alone.
• Decoding, not generating: readback refusals are typed and fail-closed; no fallback strings.
• Worktree discipline: all work in a worktree off forgejo/main; one PR per coherent phase; smoke gate in-worktree before any push.

3. Phases

Phase 0 — Worktree + record (small) — PARTIALLY DONE

• Spark PDFs copied to docs/research/.
• Adjudication doc written (spark-audit-adjudication-2026-07-18.md).
• This plan doc written.
• Create worktree off forgejo/main; run smoke + fast lanes → green baseline recorded.
• Commit the three docs + PDFs as the arc's opening record.

Phase 1 — Close the articulation seam, S1 (medium)

Wire a readback stage into the lifecycle corridor (eval-tier, off-serving):

• egress route="readback_eligible" → VocabManifold.nearest() token selection (cognitive_lifecycle.py gains a vocab consumer; today it imports no vocab, :68-76).
• Fail-closed: typed refusal when no resonant token exists — no fallback strings.
• Round-trip feedback eval ("hearing ourselves think"): re-ingest the articulated token sequence and measure phase-integrity / resonance agreement against the pre-articulation steady state. Falsifiable pass criterion fixed before implementation.
• TDD anchors: tests/test_adr_0243_cognitive_lifecycle.py, tests/test_vocab_manifold_invariants.py.
• Scope guard: within Accepted ADR-0243 §2.3 (readback rules); if design exceeds it → ADR amendment, not silent drift.

Phase 2 — Activate the learning write-path, S2 (medium)

• Compose the chiral Q_top latch (chiral_gate.py) into integrate_validated_biography (biography_wiring.py:174): charge conservation becomes a precondition of biography updates.
• First real caller: harness-driven — after ADR-0240 validation PASS, integrate holonomy; prove I-01 reboot-invariance and replay determinism in tests.
• Strictly PASS-gated + harness-driven; no idle-loop auto-integration (I-03).
• Cleanup-as-you-find: biography.py:62 bare .tobytes() → explicit <f8 coercion.
• TDD anchors: tests/test_adr_0240_biography_holonomy.py, tests/test_adr_0243_biography_wiring.py.

Phase 3 — Unify the autonomy floor, S3 (small)

• Route the lifecycle's residual check (cognitive_lifecycle.py:831) through GoldTetherMonitor so autonomy-level modulation + chiral latching govern corridor runs — one guard, not two half-guards.
• Layering direction: lifecycle → goldtether → WaveManifold (the audit's "integrate Monitor into WaveManifold" is rejected; see adjudication §2).
• Not wired into serve (A-04). Parallelizable with Phases 1–2.

Phase 4 — Generalized-lift instrument (medium-large, the crux)

Instrument before consumption (ADR-0190 lesson: build the measurement first):

• Corridor-vs-symbolic-baseline evaluation harness: problem Hamiltonians compiled from the same readers both paths use; measure solve success + refusal honesty.
• Panels: 3-domain anti-overfit panel; deductive flagship as a guard (wrong=0 must hold); real GSM8K holdout slice (never the 150 templated cases); field-reasoner reading-independence wedge.
• Anti-overfit discipline: holdout sealed; no per-case tuning; κ/γ calibration only through the existing sealed calibration evals + ratification.
• Honest-NULL protocol: if the corridor adds no eval delta, record NULL (as ADR-0246 §11 did). Truth test is eval delta, not artifact append.

Phase 5 — Handoff evidence for ruling (medium)

• Run the Phase 4 instrument through ADR-0247 ports + ADR-0248 handoff machinery (flags on in evals only) and assemble acceptance-packet evidence for ADR-0246/0247/0248.
• Deliverable: packets ready for Shay's §8 rulings. No self-Accept, no flag flips.

Phase 6 — Deferred: generate/ ("core_logos") upgrade

• Out of scope for this arc by explicit direction. Phase 4's results show where articulation quality actually binds; that evidence scopes the next arc.

4. Dependencies

• Phase 1 → Phase 4 (instrument needs the closed loop). Phases 2–3 parallel with 1.
• Phase 5 needs Phase 4's evidence. Each phase = own PR, smoke-gated, merged before the next builds on it (stacked only if necessary; merge-commit discipline).

5. Risks (honest register)

Risk	Level	Handling
Phase 4 returns NULL lift (corridor doesn't beat symbolic path yet)	High-likelihood, low-harm	Honest NULL recorded; instrument makes the question decidable; scopes Phase 6
VocabManifold coverage too sparse for meaningful readback	Medium	Fail-closed refusals are correct behavior; coverage growth is proposal-gated corpus work, not this arc
New eval suites bloat fast lane / flake in CI	Medium	Lane markers (slow/quarantine) assigned at authoring time; smoke gate at every phase
Biography wiring drifts toward autonomous mutation	Low	Harness-driven, PASS-gated only; I-03 checked in review at Phase 2 PR
Scope creep beyond Accepted ADR text	Low	Each phase names its governing ADR section; exceeding it → ADR amendment PR, not silent drift

6. Decisions to make at phase start (not blockers now)

• Phase 1: exact round-trip pass criterion (resonance-agreement threshold vs rank-agreement of top-k tokens).
• Phase 2: whether the first caller lives in evals/analogical_transfer/harness.py or evals/adr_0243_cognitive_lifecycle/ (leaning: adr_0243 harness — it already imports GoldTetherMonitor).
• Phase 4: GSM8K holdout slice size + which 3 panel domains (reuse the universal-structure trio unless evidence says otherwise).