Planning Commitment Horizon — Tasks

Before-state / evidence

• Record AI-Novel-Writing-Assistant exact implementation evidence and its soft-label counterexample in #67.
• Record Stanford Generative Agents exact implementation evidence for progressive near-term decomposition and bounded schedule rewrite.
• Audit overlap with plan.reconcile, scenario_fork, and #63 propagation debt.
• Define profile boundary: no universal chapter/volume/time horizon.

Design

• Write bilingual specification before implementation.
• Write bilingual implementation plan before implementation.
• Keep #67 independent of unmerged #65; creative-decision refs are opaque only.
• Preserve plan-vs-Canon and Settlement authority boundaries.

Stage 1 — executable contract

• Implement harness/planning_horizon.py.
• Add harness/planning_horizon.schema.json.
• Implement portable profile policy and region artifact validation/fingerprinting.
• Implement registered artifact-kind -> planning-depth admission.
• Fail closed on unknown artifact kinds.
• Implement exact-before CAS horizon transition with allowed actor classes.
• Implement one-wave dependency-evidence rebalance frontier.
• Ensure no automatic plan.reconcile, propagation-debt creation, active-plan mutation, Canon, Framework, Project, or Settlement write.

Stage 1 regressions

• soft/beat blocks chapter_plan.
• hard/chapter-detail permits chapter_plan when profile allows.
• writer cannot promote/deepen horizon.
• authorized actor can promote with exact before version/fingerprint.
• stale transition fails closed.
• unknown artifact kind fails closed.
• non-adjacent dependency with matching evidence is selected.
• adjacent region without dependency evidence is excluded.
• assumption-scoped dependency triggers only on changed-assumption intersection.
• frontier is one wave and performs no follow-up action.
• tightly plotted short-work profile can allow deep hard planning.
• discovery profile can preserve soft near-future planning.
• all artifacts/results report authority=false and model_execution=false.

Stage 2 — real runner

• Add dedicated novelforge-planning-horizon.yml.
• Compile tool and validate schema on GitHub-hosted Python 3.11.
• Run deterministic self-test and invariant assertions.
• Upload exact-head deterministic receipt artifact.
• Do not start Framework integration until dedicated workflow is green.

Stage 3 — Framework integration

• Register tool/schema/policy semantics in HARNESS_MANIFEST.yaml.
• Make dedicated workflow reusable and required by top-level normal CI.
• Integrate reusable contract checks where the latest repo architecture requires them.
• Register 011 bilingual spec/plan/tasks in documentation manifest.
• Run exact-head full NovelForge deterministic CI.
• Attribute concurrent Product/Studio baseline failures separately; do not fix them from this Core branch.

Stage 4 — semantic evaluation

• Add blind semantic capability case for meaningful far-future over-concretization.
• Add tightly plotted short-work counterexample.
• Add discovery-writing near-future-soft counterexample.
• Use a genuinely independent fingerprint-bound reviewer/model.
• If independent reviewer capability is unavailable, emit/retain PENDING_MODEL; never self-score PASS.

Promotion / rollback

• Recompare latest main before merge and preserve concurrent product work.
• Verify no Project lock, manuscript, Canon, user taste, or Settlement state changed.
• Verify #67 does not depend on #65 merge order.
• Verify disabling/removing horizon enforcement leaves existing active plans unchanged and falls back to existing planning contracts.
• Keep PR Draft until required General Craft semantic evidence is complete.