Verifiers and Judging

LRP quality labels are assigned offline. Deterministic verifiers run only inside the isolated judge worker. There is no host-process fallback.

Allowlisted verifier kinds include exact, regex, json_schema, pytest, sql_result, plugin, and none. Unknown kinds produce missing evidence, not host execution.

Seed tooling adds allowlisted tool-call plugins for presence, name, args schema, and normalized arg match. Those labels stay in separate nullable columns. They are not silently written into training quality. A tool mismatch is agreement evidence with a reference turn, not proof of task failure.

LLM judging and human audit are separate outcome classes. Verifier success, pairwise preference, and rubric scores have different meanings. Third-party judging requires operator approval for that content transfer.

Passing synthetic verifier tests does not authorize live routing activation.

Operator contracts: LRP_VERIFIERS.md. Human audit: LRP_HUMAN_JUDGE.md. Train pipeline: Train and evaluate.