implement-decision-strategy

name	implement-decision-strategy
description	Implement an explicit R0-R7 decision strategy, candidate/critic flow, retry policy, or model-only orchestration without hiding calls or assistance.

Implement decision strategy

Contract

A strategy coordinates observation, prompts, model calls, candidates, tools, verification, selection and budget. It does not mutate canonical environment state or call Stockfish unless its regime explicitly permits it.

Workflow

1. Assign the strategy an R regime and allowed H/K ceilings.
2. Define an explicit call graph and maximum calls/tokens/retries.
3. Specify prompts/output schemas and all selection policies.
4. Emit an attempt and provenance event for every call/tool/repair.
5. Separate transport retry, parse repair, legality repair, critique and best-of-N.
6. Ensure the final action is traceable to candidates and selector.
7. Add deterministic fake-model fixtures for every branch.
8. Test budget exhaustion and partial failure.
9. Verify effective assistance taint.
10. Document comparative interpretation and non-equivalent baselines.

Required questions

• Who generates candidates?
• Who evaluates them?
• Is the selector blind to candidate origin?
• Are legal moves shown before or after reasoning?
• Does any text contain SAN +/#, engine value or PV?
• Can more calls help only through a hidden oracle?

Do not call multiple samples “search” unless the system has an explicit transition/value/selection mechanism.