Calibration
MCP Atlas → OSWorld-Verified
OSWorld-Verified is estimated from MCP Atlas with a linear curve fitted on 13 models measured on both: y = 0.4874·x + 0.3766, R² = 0.33, cross-validated error 10.9 pp. It is used for 2 estimates.
| Estimated model | MCP Atlas | OSWorld-Verified | Source |
|---|---|---|---|
| LLaDA2.2-flash | 46.2% | 60.2% | estimated ± 10.9 pp, low confidence |
| Solar Open 2 | 58.2% | 66.0% | estimated ± 10.9 pp, low confidence |