Evaluator–optimizer
One produces, one scores; below threshold goes back for another pass.
Shape：A closed loop between producer and evaluator
When to use it
When quality criteria are objective or semi-objective and can be written as checks: format conformance, numbers with sources, resolvable citations, complete checklists.
When not to
Matters of taste. If the evaluator can only say "not good enough", the loop degenerates into endless rework.
How it fails
- Evaluator and producer share an origin. Same model, same biases, blind to its own blind spots — that is not an independent check.
- Criteria live in a prompt instead of the architecture. Measured on this site: when verification exists only as prompt text, the agent skips it entirely across consecutive rounds.
- Optimising to pass the evaluator rather than for quality (Goodhart).
Design notes
Make every mechanically checkable criterion an architectural gate, not a prompt request: if it does not pass, the next stage does not start. Use a different model or a different information surface for the evaluator, or the independence is fictional.
Origin
未找到可指向的首次公開實作；相關工程敘述見 orchestrator-worker 的 origin，但那是編排而非評審閉環，不混用。