porting-methodpolicy.md

MethodPolicy porting checklist

Use this for mechanism microports and analogues — declarative MethodPolicy definitions measured in the runtime matrix. Benchmark-faithful ports follow the main checklist. A named registry card alone (no MethodPolicy / matrix row) is Register a named Algorithm card.

• Declare the fidelity class up front (mechanism_microport, environment_analogue, inference_analogue, self_edit_analogue) and pin the upstream revision you traced.
• Define the method as a MethodPolicy in examples/<name>/ — frozen datasets, pure solve/propose/reward, a shared or local strategy, and a main() via examples._method_runner.standard_main.
• Put the mechanism at the standard seams: engine=Policies(selection=…, task_sampler=…, acceptance=…); a local ~15-line policy subclass where the upstream rule differs from a shipped one. Reach for aggregator_factory only when the mechanism needs state the pipeline does not keep.
• Validation lives once, in the strategy's to_diff: an unparseable proposal costs its candidate, is counted, and produces no diff — no fallback substitution anywhere, and never a gold answer in a prompt.
• Set reflective=False when artifact values are code or strict JSON (synthesised merges bypass the validator).
• Freeze evaluation splits per seed; self-generated curricula must never shape their own test set.
• Add offline tests to tests/test_candidate_methods.py (the matrix test runs every method in every scheduler without an API key) and register the builder in bench.candidate_methods.
• Add docs/algo-<name>.md in the shape every other port page has: lead blockquote, the Paper / Upstream code / Example / Domain / Layer / Fidelity table, The mechanism, Where each piece lives, Boundaries, Measured results — <domain>, Run it, and the offline-tests line.
• Add its row to the MethodPolicy table and, once measured, to all twenty-three.
• Record the port author and the upstream trace, exactly as a benchmark-faithful port does.