Numbers above reflect 100 audit passes worth of fixes. Things that used to be silently broken and now actually work:
Sharpe pass 67 was 9.6× over-reported · is now on the correct daily basis
Brier Skill pass 60 kNN no longer flips direction · is now honest
Drift PSI pass 76 5 callers couldn't reach the module · all wired up now
Covariate shift pass 79 pool too small to ever fit · now 5000-entry pool · actually fires
Per-regime calibration pass 80 regime names mismatched · now translated · stratification works
Agreement-based sizing pass 53 case mismatch silenced all checks · now active
Auto-trade entry price pass 72 opened at stale journal price · now anchors to current quote
Kelly fraction pass 74 input.fraction ignored · now respected
Webhook reliability pass 84 transient failures lost alerts · now retries 5xx/network errors
OOD detector pass 78 never decayed old samples · now EMA-Welford catches regime shifts
Service worker pass 94 JS fixes blocked behind cache · now stale-while-revalidate
Module writes pass 87, 89, 97 sample-decay/label-smoothing/self-distillation/cross-source-check all had write-storms · now cache+flush
Each row above is one diagnostic. Green = brain is doing what it's supposed to in this dimension. Yellow = warning. Red = something concrete you should investigate.
TradeTrust ≥ 75 → safe to trade automated. 60–74 → reduce size. < 60 → paper-only. < 40 → AutoPause fires.
Brier Skill > 0.10 = the brain genuinely beats baseline. < 0 = brain is anti-informative (rare, would mean something is corrupting the labels).
Annual Sharpe > 1 = professional grade. > 2 = exceptional. < 0.5 = not worth running.
Drift PSI > 0.25 = major shift, brain should retrain. > 0.10 = minor shift, monitor.
Covariate shift |AUC − 0.5| > 0.20 = the input distribution looks different from the training distribution.
Per-symbol/sector shows where the brain has measurable edge. Negative = the brain is anti-informative on that name; consider excluding.