EVALUATION / 17 SEP 2026
The endpoint is not the evidence.
Validates
Working thesis
Agents can reach useful states, skip required work and still produce misleading completion accounts. Final reports do not establish path quality or coverage.
Research questions
- 01
Separate path quality, solver quality, coverage and verified effect.
- 02
Record evaluator funding, access, method control and publication authority.
- 03
Test whether segmentation and early stopping suppress material evidence.
Connected layers
Primary sources