{"name": "maneuver-detector", "version": 1, "kind": "threshold", "features": ["delta_mean_motion"], "z_threshold": 4.0, "per_object": {"33757": {"name": "COSMOS 2251 DEB", "pairs": 343, "median_da_dt_km_day": -0.000489, "robust_sigma_km_day": 0.000319, "flag_rate": 0.02041}, "40697": {"name": "SENTINEL 2A", "pairs": 476, "median_da_dt_km_day": -0.002314, "robust_sigma_km_day": 0.010032, "flag_rate": 0.0084}, "44714": {"name": "STARLINK-1008", "pairs": 297, "median_da_dt_km_day": -0.219212, "robust_sigma_km_day": 0.233587, "flag_rate": 0.03367}, "44718": {"name": "STARLINK-1012", "pairs": 466, "median_da_dt_km_day": -0.294394, "robust_sigma_km_day": 0.23148, "flag_rate": 0.03004}}, "training": {"dataset": "Space-Track TLE history, 120 days", "objects": 4, "rows": 1582, "held_out_score": null}, "honesty": ["Unsupervised: there is no labelled maneuver ground truth here, so there is no accuracy figure to quote. The flag rate is reported instead.", "Calibrated PER OBJECT. A pooled baseline across a station-keeping constellation and an inert fragment is a baseline of nothing \u2014 the first attempt did exactly that and flagged 35% of transitions.", "An object with no calibration profile gets NO detection, which reports UNEVALUATED rather than a guess from someone else's drag regime.", "A flag means UNKNOWN, not 'manoeuvred'. It sets FR-10 to UNEVALUATED so the signal goes UNRESOLVED until fresh orbit determination arrives.", "WEAK NUMBER, VOLUNTEERED: separation between an actively station-keeping constellation and inert debris is only about 3.2% vs 2.0%. The robust z-score adapts sigma to each object, so an object that manoeuvres CONSTANTLY gets a wide baseline and its own burns stop looking anomalous relative to itself. This detector is most useful on objects that manoeuvre rarely, and least useful on the ones that manoeuvre most.", "Fixing that properly needs a physics-based drag model to predict the expected decay and flag departures from it, rather than a self-referential statistic. That is the next step, not something this achieves."], "feeds": "FR-10 \u2014 invalidates orbit determination, producing UNRESOLVED"}