Shareable benchmark table
Reference-based quality metrics
Intention-to-treat reconstructed metrics across 36 scheduled runs and 129 reference opportunities per system.
Table 02 · Reconstructed quality
Reference-based quality metrics
| Reconstructed metric | GPT-5.6 Sol · High | Grok 4.6 Build · High |
|---|---|---|
| Reference detections | 52/129 | 47/129 |
| Micro recall | 40.31% | 36.43% |
| Macro audit recall | 46.06% | 41.81% |
| Award-weighted recall | 6.81% | 4.78% |
| Closed-world precision | 50.49% | 45.63% |
| Closed-world F1 | 44.83% | 40.52% |
Codex-reconstructed manual scoring, not official API-backed EVMbench grading. Conventional accuracy is undefined.
