Anomaly detection over machine telemetry

An isolation forest over five minute windows of Sentinel's system telemetry, scored against a rolling three-sigma rule it has to beat. Working name.

detectorprecisionrecallF1 ROC AUCaverage precision
rolling mean + 3 sigma0.6920.2080.320.5860.666
isolation forest0.7190.5310.6110.6220.725
always flag (trivial)0.6161.00.762--

Held out by time, split at a gap in the archive. 211 labelled windows, positive rate 0.616. The trivial row is there because F1 on a mostly-anomalous period rewards flagging everything, and a table without it invites you to credit the model for the prevalence.

Nothing scored yet. POST a window of readings to /score.

Generated 2026-09-07T21:20:28+00:00. Recent scores are held in this process's memory and are lost on restart; there is no database and this is not a history.