| name | evaluator-optimizer |
| description | Draft an answer, judge it against a rubric, and refine on the judge's written feedback until it passes -- the generate/evaluate/optimize loop. |
| subgraph | graph.yaml |
evaluator-optimizer
Draft → judge → refine loop (benchmark T11). A reason node drafts, a
judge node scores the draft against the rubric, and on fail the
judge's rationale is re-injected into the brief so the next draft
addresses the actual objection. On pass, the graph halts.
State
- Input:
task -- what to produce.
- Outputs:
answer (the accepted draft), verdict, score,
rationale.
Use
- Standalone:
stargraph run graph.yaml (set the LM via
--lm-model/--lm-url or per-node config.model).
- As a subgraph:
kind: subgraph, spec: <path to graph.yaml>;
config.max_steps bounds the refine loop.
Tuning
Edit graph.yaml: the evaluate rubric is the whole contract -- make
it name the qualities that matter for your task.