Positioning spread: every benchmark, one model
| Benchmark | Task acc | Confidence | F₁ | Leans |
|---|---|---|---|---|
| SQuAD (factual recall) | 0.59 | 0.74 | 0.78 | +0.15 overconfident |
| MMLU-Pro (knowledge) | 0.61 | 0.92 | 0.78 | +0.31 overconfident |
| LegalBench (legal reasoning) | 0.88 | 0.82 | 0.86 | -0.06 cautious |
| MathBench (competition math) | 0.98 | 1.00 | 0.99 | +0.02 calibrated |
| OmniMath (advanced math) | 0.71 | 0.99 | 0.83 | +0.28 overconfident |
| SciCode (scientific code) | 0.59 | 0.52 | 0.64 | -0.07 cautious |
In the full cloud
gemini-3-flash-preview conditions all other model/condition points equal relative confidence and pass rate
Pairwise signal: pairs involving gemini-3-flash-preview
Match accuracy controls for the performance base-rate gap
gemini-3-flash-preview pairs
18/ 171
gemini-3-flash-preview mean tau
+0.080
All-pairs mean
+0.037
gemini-3-flash-preview p<0.05
12(66.7%)
all model pairs (observed) base-rate-matched null calibration-preserving null gemini-3-flash-preview pair (filled = p<0.05) gemini-3-flash-preview mean all-pairs mean
The four metacognitive outcomes
No curated cases for this selection yet — outcome-matrix extraction currently covers a sample of MMLU-Pro trials.