Shareable benchmark table
Reference matches and novel-valid findings
Closed-world confusion counts beside the separate expert-validity review of findings outside the reference set.
Table 03 · Finding dispositions
Reference matches and novel-valid findings
| Finding disposition | GPT-5.6 Sol · High | Grok 4.6 Build · High |
|---|---|---|
| TP / FP / FN, closed world | 52 / 51 / 77 | 47 / 56 / 82 |
| Novel-valid findings | 49 | 51 |
| Expert-validity precision | 98.06% | 96.12% |
| Conventional accuracy | N/A | N/A |
An unmatched finding is not automatically invalid. Conventional accuracy is undefined for this open-ended audit task.