Tasks 012 · AI-Native Adaptive Production Refoundation

Authority / reconciliation

• Re-bootstrap live main, PR #90 branch/base/HEAD, changed files and workflows.
• Re-read live Framework manager, session, routing, semantic-execution and self-improvement authority.
• Freeze rollback point before the refoundation continuation slice: b6f13ac97a105221f8ee78d862c4e6f02e4cf9ab.
• Separate PR #95 Godot/UI work from this SYSTEM-IMPROVE review unit.
• Classify Product/Godot/Studio/base documentation CI debt separately from candidate-owned failures.

Deterministic-overreach removal

• Remove deterministic literary class/purpose obligations from Context Assembly; retain exact-ref/stage/fingerprint safety.
• Keep context relevance/search/sufficiency model-owned through context.select.
• Remove default Reader taxonomy/HF/telemetry priming.
• Separate Blind Reader, Semantic Rule Auditor and Editor.
• Remove deterministic owner/scope → repair-depth mapping; Editor selects generation mode.
• Remove deterministic reinterpretation of character epistemic labels; retain evidence identity/story-time guards.
• Remove numeric evidence-count promotion thresholds; bind semantic promotion review instead.
• Make Author Model active preferences explicitly selected rather than automatically injected.
• Thin character/scene/realization semantic schemas.
• Keep prose telemetry optional/default-off.

Deterministic execution owners retained

• Session/checkpoint persistence and resume preflight.
• Capability/credential boundaries.
• Project/authority/write-intent isolation.
• Fingerprint, exact-source, receipt and registered-contract binding.
• Before-state/CAS/idempotency/transaction constraints.
• Stage/private-context isolation and hard budgets.
• Structural readiness/release aggregation without literary re-judgment.

Architecture / documentation

• Replace stale Spec 012 architecture with current thin-kernel / semantic-runtime owner map.
• Record current ADOPT / ADAPT / REJECT / DEFER research ledger.
• Record deterministic-overreach audit and explicit non-goals.
• Synchronize Context & Memory docs to Context Assembly v2 semantics.
• Synchronize Production Pipeline, Adaptive Learning, Harness Agent and Orchestration docs with live AI-native ownership.
• Synchronize HARNESS_MANIFEST.yaml schema IDs/semantic contract ownership.
• Remove zombie reader.production_audit authority reference.
• Restore exact harness/semantic_workers/model_contract_catalog.json authority reference required by source-integrity checks.
• Document anti-stall / stale-state execution discipline for SYSTEM-IMPROVE.

Semantic regression / ablation packets

• Add blind remote-context case (CONTEXT-REMOTE-001).
• Add irrelevant-match case (CONTEXT-IRRELEVANT-001).
• Add search-continuation case (CONTEXT-CONTINUE-001).
• Add search-stopping case (CONTEXT-STOP-001).
• Add unprimed agenda-to-dialogue Blind Reader case (READER-AGENDA-NATURAL-001).
• Preserve legitimate formal-completeness counterexamples.
• Preserve inaccessible-knowledge semantic case.
• Add plausible-inference counterexample (CHAR-INFERENCE-001).
• Preserve dynamic-planning profile cases.
• Add character-embodiment case (CHAR-EMBODIMENT-001).
• Add Reader-contamination same-candidate paired ablation.
• Add holistic-vs-decomposed Rule Auditor same-candidate paired ablation.
• Add telemetry-anchoring same-candidate paired ablation.
• Preserve deterministic unauthorized-state guards.
• Preserve deterministic stale-candidate/fingerprint guards.
• Preserve session resume stale/wrong-authority guards.
• Validate 23-case blind queue, hidden-gold isolation, three ablation same-candidate bindings and fingerprint-bound job construction in exact-head CI.
• Execute the semantic outcomes for the new regressions/ablations with an eligible independent model; currently PENDING_MODEL.

Evaluator freshness

• Re-research current official OpenAI model guidance.
• Replace stale independent-evaluator gpt-5.1 pin with current gpt-5.6 alias.
• Keep model migration separate from semantic proof: the workflow requires representative semantic execution before a semantic PASS can exist.

Deterministic verification

Evidence checkpoint f25b66749d342910bf168d27cb41cf7f66efd848 (before this ledger-only synchronization):

• Compile all 75 Python modules.
• Run candidate-owned context/semantic/quality/learning/planning self-tests.
• Build the 23-case blind semantic queue with hidden-gold isolation.
• Validate Framework version 0.8.0, compatibility and source-bundle identity in Adaptive Production Contracts.
• Confirm Semantic Contract Packs, Adaptive Production Contracts, Quality Gate Hardening and Planning Commitment Horizon workflows are green.
• Confirm AI-native semantic workflow packet preparation is green while model execution remains explicitly PENDING_MODEL.
• Re-confirm NovelForge CI hard failures are pre-existing Spec 009 bilingual/documentation-manifest debt; Python compile/version identity pass before that gate.
• Re-confirm Product Site and Studio Host Bridge already fail on base main SHA e08eb7c8cf95295221c03c55ce3e246cd07797ea; do not absorb those UI/Studio failures into PR #90.
• Confirm no downstream Project/Canon/Settlement/lock mutation and no Framework release/promotion.
• Re-run exact-head CI after this ledger-only synchronization and report it in the final handoff without creating another self-referential ledger edit.

Security / compatibility

• Search/tool access remains capability-scoped; model-owned search does not grant credentials or authority.
• Private character/creator/Reader stage boundaries remain deterministic.
• External/tool source text cannot grant Project/Canon/Framework/user-taste authority.
• Semantic results cannot self-grant write authority.
• Stale candidate/receipt/session state fails closed through exact fingerprints, before-state/CAS and resume revalidation.
• Keep Framework version 0.8.0; no downstream lock migration is required by this candidate.
• Document Context Assembly v2 migration: semantic class/purpose obligations move to model-owned selection; mechanically mandatory sources use exact refs/fingerprints.

Independent semantic gate

• Execute required blind semantic regressions/ablations with a genuinely independent model invocation bound to the exact final candidate.
• Validate independent worker identity/provenance/fingerprint result receipts.
• Current missing provider credential is represented truthfully as PENDING_MODEL / independent_provider_credential_unavailable / model_execution:false; workflow success is not converted to semantic PASS.

Release boundary

• Do not merge PR #90.
• Do not promote/release Framework behavior.
• Do not modify downstream Project locks/manuscripts/Canon/Settlement.
• Do not activate durable user taste from manager inference.

Human-review readiness

• Update PR title/body after the final exact-head CI observation so metadata describes the actual review unit and evidence.
• Required independent semantic evidence is still PENDING_MODEL; therefore READY_FOR_HUMAN_REVIEW must remain NO.