PRODUCT ADVISORY COUNCIL
Update evaluation with the task boundary
The method
Update evaluation with the task boundary
A saturated test may stop distinguishing the failures users actually experience.
When to use it
- Near-perfect benchmark scores coexist with real failures.
Ask yourself
- Tests pass. Do we still need new evaluations?
Put it into action
- Keep critical regression cases and add real new failures. Report old-baseline stability separately from new-boundary capability.
Keep these limits in mind
- Harder tests do not replace regression checks, and one high score does not solve an entire task category.