Calibration
ApprenticeBench → ExploitGym
ExploitGym is estimated from ApprenticeBench with a Michaelis–Menten curve fitted on 6 models measured on both: y = 0.7569·x / (0.50375 + x), R² = 0.77, cross-validated error 10.5 pp. It is used for 2 estimates.
| Estimated model | ApprenticeBench | ExploitGym | Source |
|---|---|---|---|
| Claude Fable 5 | 34.0% | 30.5% | estimated ± 10.5 pp, low confidence |
| Grok 4.6 | 13.0% | 15.5% | estimated ± 10.5 pp, low confidence |