EVALUATION / BENCHMARK NOTE
The Right Granularity Is Claim-Dependent
Adaptive trajectory evaluation without evidence loss at stage boundaries, milestones or early stops.
Working thesis
Evaluation resolution should be chosen by the claim being tested. Stage boundaries and early stopping can improve efficiency while hiding recoveries, alternatives and effects.
Research questions
- 01
Which evidence must cross a stage boundary intact?
- 02
When does early stopping suppress a valid recovery?
- 03
How should per-step, segment, trajectory and effect scores relate?
Connected layers