| {"name": "conformal-screening", "version": 1, "kind": "conformal_intervals", "features": [], "quantity": "miss-distance error (km) introduced by acting on an aged element set", "horizon_min": 1440, "buckets": {"0.5-1.5d": {"n": 4237, "median_km": 0.1617, "q90_km": 1.0545, "q95_km": 1.6891}, "1.5-3.0d": {"n": 145, "median_km": 0.4882, "q90_km": 3.6636, "q95_km": 5.0606}}, "coverage": 0.95, "method": "split conformal \u2014 distribution-free, finite-sample coverage. The q95 column IS the interval half-width at 95%.", "usage": "Widen a screened miss distance by the bucket matching its TLE age. If the widened interval STRADDLES a rule's limit, that rule reports UNEVALUATED rather than SATISFIED.", "training": {"dataset": "Space-Track TLE history, age-paired self-comparison", "objects": 12, "rows": 4393, "held_out_score": null, "monotonic_in_age": true}, "honesty": ["Calibrated against our own FRESHEST element set, not against independent truth. It measures how much acting on stale data would have moved our answer \u2014 which is the operationally relevant error, but it is not an absolute accuracy figure.", "Uses along-track displacement from semi-major-axis drift. Along-track dominates conjunction geometry, but it is a proxy, not a full covariance.", "We could not calibrate against operational CDMs: across 12 screened events and 37 live CDMs there were ZERO overlapping object pairs.", "Calibrated on inert objects at 600-900 km perigee, where conjunctions actually happen. Applying it to a very low or very high orbit would be extrapolation.", "Age buckets with fewer than 100 samples are NOT shipped. A TLE age with no calibrated bucket reports UNEVALUATED rather than borrowing a neighbouring bucket's number."], "feeds": "FR-01 and FR-12 \u2014 turns a point estimate into an interval, and a straddling interval into UNRESOLVED"} |