Detector accuracy & calibration.
Every detector in DebrisGuard is evaluated nightly against a held-out
slice of public data. The numbers below come straight from the
autopilot's last eval-models cycle — not a marketing
claim, not a one-time benchmark.
BSTAR-drift — historical backtest
Loading latest validation report…
Why a separate card? This is a one-shot historical backtest, not a nightly model gate — the harness
(
scripts/eval_bstar_drift_backtest.py) walks the Space-Track decay roster
backwards, scores each decayed object at multiple lead times before its actual decay,
and computes ROC-AUC + bootstrap 95 % CIs against a matched non-decay set. The
modelling cards below are for the nightly eval-models cycle.
Loading latest evaluation report…