Plan · NovelForge 7.2 Author Control + Quality Evolution

Strategy

7.2 keeps the 7.1 deterministic shell + optional semantic capability architecture. No new feature may silently invoke a model. Reader Panel, Character Integrity and revision quality passes package/validate/aggregate typed work; actual semantic execution still obeys Host Capability, Runtime Routing, and independent-session rules.

Primary loop:

observe → classify → repair owning mechanism → compare → keep/discard → regression → stop/continue

—not “ask more agents for scores.”

Per explicit user direction, implementation lands directly on main; there is no PR/feature-branch release gate. The replacement gate is stricter: exact final HEAD + deterministic CI + deterministic bundle.

Phase 0 · Baseline Audit

• Freeze the post-7.1 / Story Loom development baseline.
• Reconcile the 7.1 specification with implemented modules.
• Convert the stale 7.1 checkbox list into a historical completion ledger.
• Do not redo Story Loom/customer-facing docs.

Phase 1 · Common Quality Finding Contract

Add quality/findings.py with normalized severity/category/repair owner, candidate evidence, authority evidence, source refs, confidence, stable fingerprint, and self-test.

Phase 2 · Reader Simulation Panel

Add quality/reader_panel.py with default reading-behavior personas, single-candidate and A/B builders, per-persona order swap, validated-result aggregation, disagreement/templating/first-shown-bias diagnostics, explicit independent_gate=false, and self-test. Reuse semantic kind external_review.

Phase 3 · Durable Quality Evolution

Add quality/quality_evolution.py with SQLite runs/candidates/comparisons, exact candidate fingerprints, comparison-result consume-once, exact replay idempotency, repair-owner tracking, no-gain plateau detection, illegal winner rejection, and crash/resume tests.

Phase 4 · Context Inspector

Add harness/context_inspector.py exposing authority/stage/relevance/pin state and low-authority overlays for pin/unpin/priority/hide/invalidate-derived. Protected-authority edits become proposals. Validate pre-draft/post-draft/reviewer isolation. Never write Project Canon.

Phase 5 · Tiered + Durable Editable Memory

1. Add harness/memory_tiers.py with hot | working | archival, provenance validation, hard budgets, whole-item-or-skip, current-event/participant relevance boosts, pin-first ordering, derived authority=false, invalidation metadata, and self-test.
2. Add harness/memory_bank.py with durable SQLite entries, explicit authority classes, exact before-fingerprint edit guards, protected accepted/locked edit→proposal semantics, pin/priority controls, and Context Manifest export.
3. Proposal memory defaults to never injection stage so future/contested state cannot silently enter writer pre-draft context.
4. Free-text semantic consolidation remains a separate bounded semantic job; deterministic memory code never invents Canon summaries.

Phase 6 · Character Integrity

Add quality/character_integrity.py with bounded character snapshot + scene excerpt, agenda/knowledge/voice/relationship/spatial-task state, artifact_audit semantic job packaging, forbidden-context scan, result normalization into common findings, and self-test.

Phase 7 · State Graph Audit

Add quality/state_graph.py with scene snapshot node/edge/transition normalization, stable-attribute diff, transition-evidence binding, unexplained-change findings, before/after evidence chains, derived authority=false, and self-test. Complex narrative plausibility remains semantic work.

Phase 8 · Multi-pass Revision Orchestration

Add quality/revision_orchestrator.py:

• plan narrow continuity/character/reader/surface/research passes;
• isolate missing/failing pass execution;
• consume normalized findings and deduplicate equivalent issues;
• preserve evidence/diagnostics;
• build a repair queue by owning mechanism;
• route clustered surface failures to scene regeneration;
• route SAFE-BUT-FLAT / reader-grip failures to Reader Pressure + Scene Simulation;
• never treat a reviewer/panel pass as Canon authority.

Phase 9 · CLI / Manifest / Project SDK / CI

• Update novelforge.py to 7.2.0 and route reader-panel, quality-evolution, revision-orchestrator, context-inspect, memory-tiers, memory-bank, character-integrity and state-graph.
• Extend doctor/self-test across every 7.2 module.
• Update HARNESS_MANIFEST.yaml and SKILL* contracts.
• Update project_sdk.py so new projects default to 7.2.0 and advertise the quality-control scaffold.
• Extend reusable CI with all new deterministic self-tests while retaining all 7.1 regressions.
• Preserve deterministic bundle reproducibility and model_execution=false in normal CI.

Phase 10 · Verification

1. Run exact-HEAD compile and repository hygiene.
2. Keep all existing 7.1 contracts green.
3. Run every 7.2 self-test, including Memory Bank and Revision Orchestrator.
4. Run top-level novelforge.py self-test.
5. Verify two Framework bundle builds are byte/fingerprint identical and tamper detection remains active.
6. Assert normal CI performs no model execution.
7. Repair failures in the owning module; do not weaken the gate.
8. Only call the final exact HEAD release-ready after all required workflows are green.

Release Follow-up

• Record the exact final 7.2 commit and deterministic bundle fingerprint after green verification.
• Do not migrate the stopped legacy consumer repository in this run.
• Any future maintained consumer migration remains a separate dependency migration and must not mutate story Canon/state.

Rollback

Framework rollback base: 5e8f586b4ce0c1b90c71d0ec38064e3445daff7a. Individual 7.2 mechanisms are additionally isolated by their introducing commits and deterministic self-tests, so a failed mechanism can be reverted without inventing replacement story state.