Calibration
Toolathlon-Verified → WideResearch
WideResearch is estimated from Toolathlon-Verified with a Michaelis–Menten curve fitted on 6 models measured on both: y = 1.4847·x / (0.57158 + x), R² = 0.91, cross-validated error 3.9 pp. It is used for 5 estimates.
| Estimated model | Toolathlon-Verified | WideResearch | Source |
|---|---|---|---|
| GLM-5.3-Flash | 78.4% | 85.9% | estimated ± 3.9 pp, low confidence |
| Claude Fable 5.1 | 77.8% | 85.6% | estimated ± 3.9 pp, low confidence |
| Claude Opus 5.5 | 77.8% | 85.6% | estimated ± 3.9 pp, low confidence |
| Claude Sonnet 5.5 | 77.8% | 85.6% | estimated ± 3.9 pp, low confidence |
| Laguna S 2.1 | 49.7% | 69.1% | estimated ± 3.9 pp, medium confidence |
Fitted on
| Model | Toolathlon-Verified | WideResearch |
|---|---|---|
| dots3-note Preview | 55.6% | 78.9% |
| Hy4 preview | 74.1% | 83.9% |
| Ornith-1.5-35B-A3B | 48.7% | 67.8% |
| Ornith-1.5-397B | 71.2% | 80.8% |
| Ornith-1.5-9B | 41.2% | 59.5% |
| Qwen3.8 Max | 72.5% | 81.9% |