Feedback-to-plan: choose, build, challenge and revise

12 September 2026 · Experiment 05 · higher-autonomy Gimmer

Lawrence authorised the next substantive step and supplied eleven additional source records. The records were located and consulted; five lifecycle captures are excerpts of one paper. Their methods changed the immediate priority from a reusable module-composition experiment to explicit feedback, selective rework and observed-state response policies.

Construction

The new engine explores native repeated Petri transitions within a declared one-to-eight-event horizon. It distinguishes planner choices from uncontrolled outcomes, reports all winning choices, constructs a policy minimizing worst sequential duration, and checks that policy across all its environment branches. A realised successful history is separately expanded through the existing token-ancestry and typed-composition machinery into an occurrence-specific work plan and resource-checked schedule.

The refuge examples distinguish selective and blanket rework, current evidence and historical approval, retained independent work and consumed fallback resources. The permit pair exposes the gap between a successful plan for each separately known scenario and one decision policy that cannot know the future. The repeatable-review example retains adverse outcomes at every depth.

Independent challenge

An independent oracle enumerates complete policies over observed-state/horizon keys before unfolding their environmental outcome paths. It compares 81 tiny game topologies across four horizons, checks the returned policy at every state, and independently replays token supplies and schedule demand. This provides a separate construction against which to challenge the production AND/OR solver.

Review found and corrected a substantive model gap: drawings could previously be renewed before the structural analysis they depend on. The corrected mechanism requires current analysis before redrawing after major feedback, while a drawing-only correction preserves valid analysis. A regression exercises both paths.

The 108 retained model tests and 30 new focused/oracle tests pass. The ordinary interface passes 40 checks through model selection, feedback, repeated work, policy comparison, completed-history inspection, save/reopen/correction and desktop/phone use. This includes the collection routes, earlier experiments, canonical inherited baseline, method and run readers. The retained Marking-to-Plan and Process Contract browser journeys also pass all 37 and 39 checks. The final private run receipt records the release revision, completed test counts, live journeys and exact served-source verification.

Scope and continuation

This is a bounded method experiment using fictional assumptions. It supplies agent-operated mathematical and interface evidence, not construction engineering approval, calibrated design predictions or independent human-use validation. Existing experiments remain available in the same collection. The separate composition frontier is retained as an open research question.

Use the experiment · Read the method and sources · Return to the collection