Results
Seeded benchmark, produced by chronotrace eval. The corpus is written for this project rather than mined from the wild, so this is a controlled comparison on identical inputs — a directional effect, not a population estimate.
Arm C
15 cases · provider bedrock
- Repaired and verified
- 86%
- Wrongly patched a non-race
- 0%
- Refused correctly when it should
- 100%
- Timing band-aids introduced
- 0%
- Proposed orderings that reproduced the failure
- 82%
- Measured overhead
- 0.1815 msno fixed delay introduced
Verification strength: no patch reached verification ×8, forced_harmless ×6, not verified ×1