Calibrated Panel Quality Validation on rag-mini-wikipedia clean regime
85.8EQ (%)Claude Sonnet 4.6
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| Claude Sonnet 4.6Retrieval depth (top-k)=15, Multiplicity=1.148, Byte reduction=14.13%, n=400, Number of judges=52026.05 | 85.8 | 11.5 | 1.75 | 2 | 1 | 1.4 | — | |
| Gemini 2.5 FlashRetrieval depth (top-k)=15, Multiplicity=1.148, Byte reduction=14.13%, n=400, Number of judges=52026.05 | 85.8 | 11.3 | 2.75 | 0.3 | 6 | 3.24 | — | |
| GPT-5.1Retrieval depth (top-k)=15, Multiplicity=1.148, Byte reduction=14.13%, n=400, Number of judges=52026.05 | 84 | 14.3 | 1.5 | 0.3 | 4 | 2.55 | — | |
| Llama 3.3 70BRetrieval depth (top-k)=15, Multiplicity=1.148, Byte reduction=14.13%, n=400, Number of judges=52026.05 | 76.8 | 18.8 | 4.5 | 1 | 6 | 3.25 | — |