Calibration
ERQA → RealWorldQA
RealWorldQA is estimated from ERQA with a Hill curve fitted on 6 models measured on both: y = 0.0000 + (0.9034 − 0.0000)·x^6.00 / (0.40223^6.00 + x^6.00), R² = 0.91, cross-validated error 1.1 pp. It is used for 5 estimates.
| Estimated model | ERQA | RealWorldQA | Source |
|---|---|---|---|
| Claude Opus 4.6 | 51.6% | 73.8% | estimated ± 1.1 pp, low confidence |
| Gemini 3.1 Pro | 69.4% | 87.0% | estimated ± 1.1 pp, medium confidence |
| GPT-5.4 | 65.4% | 85.7% | estimated ± 1.1 pp, medium confidence |
| Grok 4.20 | 54.1% | 77.3% | estimated ± 1.1 pp, low confidence |
| Muse Spark | 64.7% | 85.4% | estimated ± 1.1 pp, medium confidence |