Plan 012 · AI-Native Adaptive Production Refoundation

Goal

Finish the existing PR #90 review unit as a coherent thin-kernel / model-owned semantic-runtime candidate. Do not create parallel owners, do not merge/release, and do not mutate downstream Projects.

Live reconciliation

Live bootstrap for this refoundation is always resolved from the current PR/HEAD, not from the SHAs written in older chat or historical plan text. The rollback checkpoint immediately before this synchronization slice was b6f13ac97a105221f8ee78d862c4e6f02e4cf9ab; every later consequential write must re-read PR #90 before-state.

Current owner decisions:

• KEEP Session Runtime, Control Plane, exact identity/fingerprint, CAS, receipts, stage isolation and hard budgets.
• THIN Context Assembly, Writer realization and repair routing to machine-required fields only.
• MIGRATE semantic relevance/search, Reader experience, rule applicability, repair depth, planning quality, character reasoning and feedback interpretation to model contracts.
• OPTIONAL_TOOL prose telemetry.
• MERGE_WITH_EXISTING_OWNER rather than invent second stores/readers/simulators/releases.
• DEFER new languages/runtimes/dependencies until a concrete owner and measurable need exists.

Phase A · Authority and stale-state cleanup

1. Re-bootstrap PR #90, main base, changed files, workflows and exact Framework authority.
2. Treat PR body/old plan SHAs as historical notes only.
3. Synchronize HARNESS_MANIFEST.yaml schema IDs and contract names with live code.
4. Remove stale assertions that Python chooses literary repair depth or semantic context obligations.
5. Keep Framework version 0.8.0; no release/promotion.

Phase B · Semantic-role separation

1. Keep reader.engagement_audit as the production Blind Reader.
2. Keep quality.semantic_rule_audit as a separate hard-rule semantic role.
3. Keep quality.production_review as the genuinely independent holistic gate when enabled.
4. Keep editor.repair_spec as the semantic repair owner.
5. quality/repair_policy.py receives Editor-selected generation mode and enforces only the resulting writer-information boundary.

Phase C · Agent-owned search/context

1. context.select owns missing-information diagnosis, query formation, relevance, reformulation, continuation and stopping.
2. context_inspector.py owns eligibility/stage/protected-edit mechanics, never relevance.
3. context_assembly.py v2 validates only exact selected refs, exact higher-authority required refs, source fingerprints, stage safety and private-state boundaries.
4. memory_tiers.py may pack whole selected blocks under hard budgets; it must not convert priority/top-k into literary truth.
5. Remove documentation that says a deterministic class/purpose obligation proves semantic sufficiency.

Phase D · Planning / simulation / realization

1. Preserve commitment strength/depth/CAS/fingerprint mechanics in planning_horizon.py.
2. Make explicit that the Planner, not the runtime, chooses useful depth/uncertainty.
3. Keep private character state → action → scene collision as semantic work.
4. Keep runtime checks to evidence identity, story-time eligibility, permissions and visibility.
5. Keep writer projection compact; do not re-expand a Realization Sheet.

Phase E · Learning

1. Semantic learning.preference_interpret interprets feedback.
2. LearningStore/Author Model persist evidence/hypotheses under CAS and authority.
3. active remains eligibility only; manager/model selects relevant hypothesis IDs.
4. Promotion Gate verifies a bound semantic review plus durable-write prerequisites; no numeric evidence-count literary threshold.

Phase F · Ablation and regression coverage

Add/retain blind semantic families for:

• remote context outside a recent horizon;
• irrelevant similarity match;
• search continuation;
• search stopping;
• agenda-to-dialogue artificiality without exposing expected HF code;
• legitimate formal completeness;
• inaccessible knowledge and plausible inference;
• dynamic planning profiles;
• character embodiment;
• Reader contamination (unprimed vs taxonomy-primed);
• rule audit (holistic vs decomposed mandatory rules);
• telemetry anchoring (not preloaded vs preloaded);
• deterministic unauthorized-state rejection;
• deterministic stale-candidate rejection;
• long-horizon resume / stale checkpoint authority revalidation.

Pair semantic ablations in evals/ai_native_ablation_manifest.json. The ablation manifest must bind the same candidate text/authority across conditions. Deterministic CI validates packet construction; semantic outcomes require a real model worker and remain PENDING_MODEL when that capability is absent.

Phase G · Current evaluator freshness

The independent semantic workflow must not hard-pin a superseded model indefinitely. Current primary-source research identifies the OpenAI gpt-5.6 alias as the current complex-production baseline, so the workflow should use that alias rather than the older gpt-5.1 pin. This is an evaluator-infrastructure freshness change, not semantic evidence by itself.

Future model changes must follow the same rule: research current official guidance, run representative evals, and update only with evidence. Model popularity alone is not authority.

Phase H · Deterministic verification

Run all candidate-owned focused checks, including:

• Python compile;
• context inspector / context assembly;
• semantic router/catalog/reference integrity;
• repair policy, production readiness/release, telemetry;
• promotion gate / Author Model;
• planning horizon;
• bundle/version identity and deterministic double-build where the workflow already owns it;
• blind queue construction and hidden-gold guards;
• exact-candidate receipt/fingerprint/independence validation.

Do not add Python tests that claim to prove prose quality.

Phase I · CI anti-stall and stale optimization

Adopt the useful execution discipline from the UI completion contract without importing UI scope:

• every poll/build/workflow observation is bounded;
• no WAITING state may persist without new evidence;
• if the same workflow remains pending across two observation cycles, inspect jobs/logs rather than sleeping again;
• every consequential write revalidates PR branch, HEAD, base and exact before-state;
• each failed job is classified as candidate-owned, pre-existing/base, external-capability, or transport/configuration;
• retry only after the cause changes; never retry identical failures indefinitely.

Phase J · Security / compatibility

Verify:

• model can choose searches only inside allowed capability scope;
• source text cannot grant permissions or redefine authority;
• credentials stay outside semantic context;
• Reader/private-character/manager information boundaries remain intact;
• semantic results cannot self-grant Project/Canon/Framework/user-taste authority;
• stale candidate/receipt/session state fails closed;
• no downstream lock/manuscript/Canon/Settlement changes occur.

Exact CI classification policy

Pre-existing Product/Godot or Studio failures are not evidence that this candidate failed. They must still be reported, but SYSTEM-IMPROVE only repairs them if they directly block this architecture work. Candidate-owned deterministic workflows must be green at the final HEAD.

A workflow that records PENDING_MODEL is functioning correctly but does not provide semantic PASS.

Rollback

• rollback base for this synchronization slice: b6f13ac97a105221f8ee78d862c4e6f02e4cf9ab;
• each documentation/authority/eval slice lands as a reviewable commit;
• no force-push/history rewrite;
• downstream consumers remain pinned, so reverting PR #90 cannot change their live Framework authority.

Stop condition

Normal completion requires all executable deterministic, documentation, ablation-packet, CI, security and compatibility work to be complete and required independent semantic evidence to exist for the exact final candidate. If independent model capability is unavailable, complete everything else and stop only with the explicit external blocker PENDING_MODEL.