Positioning spread: every benchmark, one model
| Benchmark | Task acc | Confidence | F₁ | Leans |
|---|---|---|---|---|
| SQuAD (factual recall) | 0.45 | 0.64 | 0.68 | +0.19 overconfident |
| MMLU-Pro (knowledge) | 0.45 | 0.99 | 0.62 | +0.54 overconfident |
| LegalBench (legal reasoning) | 0.85 | 0.70 | 0.73 | -0.15 cautious |
| MathBench (competition math) | 0.89 | 1.00 | 0.95 | +0.10 overconfident |
| OmniMath (advanced math) | 0.47 | 0.98 | 0.64 | +0.51 overconfident |
| SciCode (scientific code) | 0.51 | 0.95 | 0.65 | +0.44 overconfident |
In the full cloud
gemini-2.0-flash-001 conditions all other model/condition points equal relative confidence and pass rate
Pairwise signal: pairs involving gemini-2.0-flash-001
Match accuracy controls for the performance base-rate gap
gemini-2.0-flash-001 pairs
18/ 171
gemini-2.0-flash-001 mean tau
+0.027
All-pairs mean
+0.037
gemini-2.0-flash-001 p<0.05
8(44.4%)
all model pairs (observed) base-rate-matched null calibration-preserving null gemini-2.0-flash-001 pair (filled = p<0.05) gemini-2.0-flash-001 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.