Cost And Proof Freshness
Expensive Cautilus eval and improve evidence should be shown honestly. Reader-facing reports should distinguish selected evidence, prepared evidence, stale evidence, and newly executed proof.
Key: rule.cost-and-proof-freshness.
Where To Check This
- User-facing attachments: Behavior Evaluation, Bounded Improvement
- Maintainer evidence routes: Evaluation Surfaces And Runners, Improvement Loop, Scenario History And Proposal Normalization
Evidence State
The held-out improve cycle is now proven live on the dev/skill surface: npm run proof:improve:live runs a real bounded improve loop that recovers a held-out scenario it was never tuned on, with the operator-witnessed capture under fixtures/eval/dev/skill/improve/live/ replayed deterministically and projected by Bounded Improvement.