SHADOW-MODE EVALUATION · SYNTHETIC DATA

ML predictions

Trained in your browserBrowser demo · no AWS calls
PILLARS 1 AND 5 · STRATEGY AND COMPONENT HEALTH

Models you can inspect, tested the way the pilot would.

Two models train on synthetic sessions here and are scored on later, held-out races. The models are real; the data are generated, so results are optimistic. Real telemetry is noisier, which is exactly what the pilot’s shadow mode measures.

FAULTS FLAGGED BEFORE FAILURE—held-out races
FALSE ALERTS / RACE WEEKEND—an alert needs two laps above threshold
MEDIAN WARNING TIME—before the failure lap
PILOT GATE (APPENDIX F)—at least 80% flagged, at most 1 false alert per weekend
01 / PREDICT

Component early warning

HELD-OUT RACES ONLY
Fault score Alert threshold● valid alert · shaded: fault developing
—
—
02 / TUNE

Alert threshold

EDITABLE

Lower catches more faults and raises false alerts. The default is tuned on the 24 training weekends so that false alerts stay at or below one per weekend, then scored on the 12 held-out weekends.

What the model learned

    Standardised logistic-regression weights. Positive means the signal raises risk. An engineer can read this; that is the point of starting simple.

    STRATEGY INPUT

    Tyre-degradation prediction.

    A regression model trained on 360 synthetic stints, scored on the next 120. It predicts pace loss per lap with an uncertainty band and suggests a stop age. It feeds the strategy optimiser; the engineer still decides.

    PREDICTED PACE LOSS NOW——
    HELD-OUT ERROR (MAE)—per lap, on 120 later stints
    MODEL-SUGGESTED STOP——
    MODELRidge regressionage, age², temperature, age × temperature
    03 / PREDICT

    Pace loss versus tyre age

    90% BAND
    Grey dots: held-out stints near the chosen temperature · red line: current tyre age · blue line: stop trigger
    04 / EXPLORE

    Race conditions

    EDITABLE

    Hotter tracks wear tyres faster, so the suggested stop moves earlier. Synthetic data; not a validated tyre model.