Healthy distribution:
- Spread across [0.20, 0.80] with some predictions in both confident tails
- Mean near 0.50 (no strong systemic long/short bias)
- Std around 0.10-0.20 (the brain is differentiating)
- Confident% > 20% (model has some strong opinions)
- Muddy% < 30% (not just outputting 50% for everything)
Unhealthy patterns:
- All in [0.45, 0.55]: the brain is uncertain about everything. Probably under-trained or the features are too weak.
- All at one end (e.g. 0.70+): the brain is biased — probably wrong calibration or symbol bias overshoot.
- Bimodal at 0.0 and 1.0: overconfident model. Calibration is needed (or label smoothing should be turned up).