Skip to content
NovelForge DocsReference

Planning Commitment Horizon — Implementation Plan

Implement #67 as a small deterministic planning-admission and rebalance-frontier contract layered on existing NovelForge planning primitives. Do not introduce a second active-plan database or any Canon authority.

Implement #67 as a small deterministic planning-admission and rebalance-frontier contract layered on existing NovelForge planning primitives. Do not introduce a second active-plan database or any Canon authority.

  • Base the branch on an exact current main commit.
  • Treat current plan.reconcile, scenario_fork, #63 propagation debt, Context, Canon/Settlement, and semantic-worker contracts as existing owners.
  • Keep #65 an optional opaque-reference integration only; this branch must build and run if #65 is not merged.

Phase 1 — portable region/policy contract

Section titled “Phase 1 — portable region/policy contract”

Create harness/planning_horizon.py and a machine schema.

The deterministic contract should support:

  1. canonical fingerprinting of policy/region artifacts;
  2. small fixed planning-depth registry and known artifact-kind mapping;
  3. region creation/validation with open|soft|hard strength and a depth ceiling no greater than the selected profile policy allows;
  4. realization admission for a registered artifact kind;
  5. explicit CAS horizon transition with allowed actor classes and evidence refs;
  6. one-wave rebalance-frontier selection from explicit dependency evidence;
  7. authority=false on every artifact/result and no model execution.

Self-test at minimum:

  • soft/beat blocks chapter detail;
  • hard/chapter-detail permits chapter plan when profile permits;
  • unknown artifact kind fails closed;
  • writer cannot promote;
  • authorized planner/user can promote with exact before state;
  • stale version/fingerprint fails;
  • identical retry is deterministic and does not create hidden write authority;
  • non-adjacent linked dependency is selected;
  • adjacent unlinked dependency is excluded;
  • assumption-scoped dependency only triggers when changed assumptions intersect;
  • frontier result performs no reconcile/debt/Canon operation;
  • tight-short and discovery profile counterexamples are both representable.

Add a dedicated workflow that:

  • compiles the tool;
  • validates schema JSON;
  • runs self-test;
  • asserts authority/no-model/no-auto-action invariants;
  • uploads the deterministic receipt.

Do not modify HARNESS/docs/normal CI until this workflow passes on the real GitHub runner against the current repository.

After dedicated CI is green:

  • register the tool/schema in HARNESS_MANIFEST.yaml;
  • make the dedicated workflow reusable and a normal-CI job;
  • add any reusable-contract invocation needed by the current repository architecture;
  • register bilingual spec/plan/tasks in documentation governance;
  • preserve all concurrently developed Product/Studio files from current main.

Phase 5 — semantic capability/counterexample eval

Section titled “Phase 5 — semantic capability/counterexample eval”

Use existing NovelForge blind eval infrastructure. Add profile-boundary cases such as:

  • capability: a serialized novel has only a soft far-future arc role, but a planner tries to generate detailed chapter plans several arcs ahead. Expected route: block/escalate for horizon promotion rather than silently concretize.
  • counterexample: a short mystery has an explicit user requirement for a fully plotted 12-chapter structure before drafting. Expected route: allow a profile that hard-plans the whole work rather than forcing artificial softness.
  • counterexample: a discovery-writing profile deliberately keeps even the next chapter’s later scenes soft. Expected route: preserve soft planning rather than automatically promoting because the material is near-term.

Missing independent model capability must remain PENDING_MODEL, not be converted into self-scored PASS.

Before merge/promotion:

  • verify exact-head deterministic workflows;
  • verify semantic evidence or leave PR Draft if semantic evidence remains pending;
  • confirm no Project/Canon/manuscript changes;
  • confirm no hidden dependency on #65;
  • confirm rollback is deletion/disable of horizon enforcement without data migration;
  • compare latest main again and preserve concurrent Product/Studio work.