| {"name": "density-uncertainty", "version": 1, "kind": "linear", "features": ["kp"], "intercept": 0.425714, "coefficients": [0.037019], "bins": [{"kp_lo": 0, "kp_hi": 1, "n": 867, "median_decay_km_day": -0.293801, "relative_sigma": 0.5098}, {"kp_lo": 1, "kp_hi": 2, "n": 1398, "median_decay_km_day": -0.291681, "relative_sigma": 0.4867}, {"kp_lo": 2, "kp_hi": 3, "n": 1047, "median_decay_km_day": -0.318938, "relative_sigma": 0.4952}, {"kp_lo": 3, "kp_hi": 4, "n": 555, "median_decay_km_day": -0.361624, "relative_sigma": 0.5071}, {"kp_lo": 4, "kp_hi": 5, "n": 218, "median_decay_km_day": -0.34226, "relative_sigma": 0.5202}, {"kp_lo": 5, "kp_hi": 9.1, "n": 117, "median_decay_km_day": -0.368602, "relative_sigma": 0.759}], "shape": "threshold, not ramp \u2014 roughly flat below Kp 5, then a step of about 1.5x. Consumers should prefer the bins over the linear fit.", "quiet_sigma": 0.5038, "storm_sigma": 0.759, "training": {"dataset": "Space-Track TLE history (drag-only objects) x GFZ Potsdam Kp", "kp_source": "https://kp.gfz.de/app/files/Kp_ap_since_1932.txt (CC BY 4.0)", "rows": 4202, "objects_used": ["ARIANE 44L+ DEB (SPELDA)", "PEGASUS R/B", "DELTA 2 R/B(2)", "COSMOS 2251 DEB", "CZ-3C R/B", "ARIANE 5 R/B", "ARIANE 5 R/B", "ARIANE 5 R/B", "FALCON 9 R/B", "CZ-7A R/B", "CZ-5 R/B", "KZ-11 R/B", "ELECTRON R/B"], "objects_excluded": ["SENTINEL 2A", "STARLINK-1008", "STARLINK-1012"], "held_out_score": null, "slope_positive": true}, "honesty": ["Calibrated on a SMALL set of drag-only objects over 120 days. It is a measurement of this sample, not a global density model.", "Station-keeping objects are excluded because their decay rate includes thruster firings \u2014 including them would measure the operator, not the atmosphere.", "Samples are NORMALISED PER OBJECT before pooling. Pooling raw decay rates across objects with perigees from 186 to 302 km measures the altitude spread, not the atmosphere \u2014 the first version did exactly that and returned the MAD-to-sigma constant in every bin.", "The high-Kp bins are thin because the 150-day window was geomagnetically quiet. Extrapolation to storm conditions is exactly that: an extrapolation.", "ML-HASDM (Licata et al., Space Weather 2022; ~11% error with calibrated uncertainty) is the correct upgrade and slots into the same field.", "If the fitted slope were negative, this would be reported as a FAILED calibration rather than shipped.", "FINDING: the hardcoded model this replaces used 0.15 at quiet conditions. The measured value is about 0.49 \u2014 more than 3x more uncertain than assumed. The hardcode was optimistic where it mattered least and roughly right where it mattered most.", "FINDING: median decay itself rises about 25% from Kp 0-1 to Kp 5+, so a storm both accelerates decay AND widens its uncertainty."], "feeds": "the Monte Carlo re-entry footprint dispersion, and FR-21 model validity"} |