Routing review
What did this model add, and is it worth routing this kind of task to it again?
Compare two completed runs side by side. Validated evidence, subjective scores, warnings and failures stay separate on purpose — there is no single combined score to hide behind.
- Validated evidence — checks, time, tokens
- Derived — computed from evidence
- Subjective — reviewer and your scores
- Warnings — reported issues
- Failure — no validated artifact
Swap with the button or press X when focus isn’t in a text field. Your scores travel with each model.
Head-to-head:
Each signal is read separately. Underlined value = the side that leads on that row.
Decision record
What tipped the decision? Mention evidence vs. judgment explicitly. Not saved automatically — reloading the page clears scores and notes.