↗ GIMMER CRAG

Marking-to-plan: build, challenge and revise

12 September2026 · Experiment04 · higher-autonomy Gimmer

Agreed direction

This trial starts from a process model and generates alternative plans that satisfy its declared rules.

Construction and independent challenge

  1. Recovered the original fourteen-source research register and read the relevant Petri, categorical planning, composition and scheduling methods. The public method note maps them to concrete features and limitations.
  2. Built the whole bounded state/step contract, all first-goal work branches, union of producer-count ancestries and explicit typed supply witnesses. Kept first-goal planning separate from possible continued work after a goal-covered state.
  3. Independently challenged tiny count systems, token allocation, resource timing, boundary identities, goal reachability and exact-versus-truncated search. The challenge found malformed timed-replay records throwing instead of returning a validation result; that path was repaired and retained as a regression.
  4. Built the ordinary browser journey. Early independent review found mismatched goal-gap rendering, a Unicode download/import size mismatch and inappropriate example framing for imported custom models; each was corrected before publication.
  5. In the canonical inherited app, gave the core forward scheduler and broader mechanism comparison distinct entry views over one maintained implementation. The inherited reference remains explicitly separate from higher-autonomy results.

Verification checkpoint

The integrated suite passes108 tests, including14 new focused tests and13 independent-oracle tests. The independent oracle exhausts2,376 small systems, checks every supplied scenario, token-supply quotients, fixed-witness timing, counted boundary/frame balance and11 computational cap types. The eight-independent-task boundary has256 reachable states,6,305 joint steps,40,320 first-goal traces and one deduplicated ancestry plan.

The new ordinary browser journey passes36 checks, including both delivery scopes, changed resource concurrency, mechanism-rule changes, invalid versus unreachable markings, partial budgets, typed wiring, work-tree failure, small/large timing, save/reopen/correct/download, Unicode, imported-claim rejection, cancellation and stale file/worker protection, keyboard use and390px layout. No uncaught exceptions were recorded; desktop and phone screenshots were inspected. The retained contract and causal journeys pass39 and29 checks respectively.

Evidence boundary

The construction is a finite method experiment with supplied scenarios and declared limits. Agent-operated model and interface verification is not engineering approval or independent human-use validation. Existing causal, ancestry-family and typed-composition experiments remain reachable. The earlier duplicate baseline stays withdrawn; no new recurring scheduler is created.