Training & audits

These pages cover continuous learning and the fairness audits. None of them place trades.

Training /training

Status and control for the three continuous-learning jobs. A status panel shows the running job (progress bar) or Idle + last summary. Config cards show the training symbol count, forecaster, device, confidence gate, target hit-rate, and min support. Then the three jobs:

JobCadenceWhat it does
RecalibratenightlySets the confidence gate from fresh out-of-sample data.
Wheel optimizerweeklySuggests a put delta × DTE (CPU-friendly).
Fine-tuneweeklyRolls a Kronos fine-tune (heavy GPU); only adopted if it out-calibrates the incumbent.

Each shows an on/off pill and its last-run metric. A Run now button appears only for the platform owner; other accounts inherit the recalibrated settings automatically. A calibration-curve table (confidence bucket → hit-rate → samples) and a training activity feed round out the page.

Bias audit /bias admin

Quantifies directional skew so "no bias" is measured, not claimed. Three sections: (1) Directional balance from your own history — the council approval gap long vs short (green if within 8 pts); (2) a Council mirror test that feeds four matched bull/bear pairs and reports "symmetric" vs "±X pts toward longs/shorts"; (3) Forecaster calibration parity — do up-calls and down-calls hit their confidence equally.

Forecast lab /forecast_lab owner

Platform-wide forecaster experimentation (owner only): toggle Chronos-2 covariate signals, run an A/B hindcast (which covariates actually improve accuracy, paired + significance-tested), and run an option-P&L sniping hindcast (simulate buying an ATM option in the forecast direction to test whether the signal is tradeable). Both are accuracy tools — a negative result is a hard no.