How to read this
Left — map (metres from the circling centroid): each of the 14 drone rotations gives one bearing ray from where the drone was. Orange dotted = current estimator (phase-contrast argmax on the camera axis). Purple = antenna-model fit (Σ/Δ squint removed, calibrated CI). Shaded ribbons = 95%/99% credible regions after fusing all rotations.
Right — the same bearings without map clutter. One row per rotation, bar = 95% CI. Two things to see: purple points scatter about half as much as orange, and the orange CIs are erratic — from an implausible ±0.06° (C2R1) to ±25° — because they bootstrap a single smoothed curve; the purple CIs are calibrated and consistent.
Result at 95%
old model cross-range width 327 m 280 m region area 1.09 km² 0.91 km² bearing scatter σ 3.7° 2.3°
Range is the honest weakness — for both methods. All rays leave a 297 m circling patch pointing the same way (bearings 11–27°), so the credible region is a 5–6 km ribbon along the rays; the two shipped fusers already disagree by 764 m in range. Better bearings narrow the ribbon, they cannot shorten it — only wider flight geometry can.
Squint this flight is only -0.4°, so the fused peaks nearly coincide; in the calibrated simulation the current estimator carried a +2.75° squint bias where the model's advantage is much larger. The model also recovers the 2 C1 rotations the current pipeline drops.