| name | topology-experiment-evaluator |
| description | Evaluate meta-runtime topology experiments using explicit evidence and promotion criteria. |
| version | 1.0.0 |
| metadata | {"sloppy":{"tags":["runtime","experiments","evaluation"],"category":"runtime"}} |
Topology Experiment Evaluator
When To Use
Use this skill when a topology proposal has an experiment and needs evidence before promotion or rollback.
Procedure
- Read the experiment, linked proposal, criteria, and recent route events.
- Separate direct evidence from inference. Direct evidence includes dispatch status, child completion state, approval outcomes, and user-visible failures.
- Score only the criteria stated on the experiment. If criteria are vague, record that gap instead of inventing a metric.
- Use
record_evaluation to attach evidence and a clear pass/fail or numeric score. Include the route/event ids, child session ids, approval ids, or user-visible failures that justify the score.
- Promote only when your evaluation concludes the recorded evidence satisfies the criteria. Pass the evaluation id to
promote_experiment; the provider records that evidence but does not score the criteria for you.
- Roll back with an explicit rollback proposal when a promoted variant regresses.
Pitfalls
- Do not treat a quiet event log as proof of success.
- Do not promote a route because it works for one unrelated envelope.
- Do not compare variants without recording the envelope mix and time window.
Verification
Confirm the evaluation appears under /evaluations, promoted experiments record promotionEvaluationId, and promotion or rollback uses ordinary proposal affordances.