Positioning spread
| Benchmark | Task acc | Confidence | F₁ | Leans |
|---|---|---|---|---|
| SQuAD (factual recall) | 0.48 | 0.40 | 0.72 | -0.08 cautious |
| MMLU-Pro (knowledge) | 0.65 | 0.67 | 0.74 | +0.03 calibrated |
| LegalBench (legal reasoning) | 0.86 | 0.64 | 0.74 | -0.23 cautious |
| MathBench (competition math) | 0.95 | 1.00 | 0.97 | +0.05 calibrated |
| OmniMath (advanced math) | 0.58 | 0.74 | 0.75 | +0.16 overconfident |
| SciCode (scientific code) | 0.56 | 0.65 | 0.73 | +0.09 overconfident |
In the full cloud
Claude Haiku 4.5 conditions all other model/condition points equal relative confidence and pass rate
Pairwise footprint
Match accuracy controls for the performance base-rate gap
Claude Haiku 4.5 pairs
18/ 171
Claude Haiku 4.5 mean tau
+0.058
All-pairs mean
+0.037
Claude Haiku 4.5 p<0.05
9(50%)
all model pairs (observed) base-rate-matched null calibration-preserving null Claude Haiku 4.5 pair (filled = p<0.05) Claude Haiku 4.5 mean all-pairs mean