🏥 Model Health

Every settled pick pairs the generator's stat-scale projection (expected Ks, walks, runs, points…) with the real box-score number the grader settled it with. Bias > 0 = the model projects too high; the shrunk correction is what a generator would subtract to fix it. Grows daily from the free feeds — fully independent of any paid odds key.

Each component scores 0-100. Tiered weights aggregate into a single grade:

• Calibration freshness (15%): How recent are the per-sport bias shifts in self_training_*.json?
• Hit rate trend (20%): 7-day rolling hit rate on settled props (target >70%)
• Brier trend (20%): 7-day rolling Brier score (target <0.15)
• Pipeline audit (15%): Green/yellow/red ratio across all data artifacts
• Data health (20%): Live cross-checks against MLB / ESPN / Daum upstream APIs
• Edge count (10%): Number of actionable picks surfaced today

Tiers: ELITE ≥85 · HEALTHY ≥70 · OK ≥55 · DEGRADED ≥40 · BROKEN <40