Evaluator Optimizer

One model generates, another evaluates, and the loop stops when the rubric passes.

Active scenario: loop. A generator drafts, an evaluator scores against a rubric, and feedback drives revision.

Metrics: quality signal: high; token spend: medium; stop clarity: high.

Nodes: Request, user: The task arrives with an explicit success criterion. Generator, worker: The optimizer proposes the next candidate answer or patch. Evaluator, review: A separate pass checks the candidate against concrete criteria. Accepted, state: The loop exits only when the score clears the threshold or budget ends.