Satellites in low Earth orbit avoid one another through small orbital adjustments. Each adjustment has an unaccounted consequence: for several hours, the public catalogue used for conjunction screening inaccurately represents the manoeuvring satellite. We quantify this inaccuracy using 620,275 public element sets covering 1,800 satellites over 120 days. manoeuvred satellite has 2.23× greater prediction error than a non-manoeuvring satellite, almost entirely along the direction of travel. Screening real orbits against stale entries produces 120 false conjunction alerts and misses 111 genuine alerts, relative to 136 genuine conjunctions. Decomposition shows that injected error increases with both manoeuvre magnitude and notification delay. Prior work addresses the former; we show that the latter dominates. For a median manoeuvre, the 3.92 h delay contributes 4.73 km of error against a 5 km alert threshold, whereas immediate notification contributes almost none. We therefore propose publishing a manoeuvre when it is planned, using the command link through which it is executed; every manoeuvrable object already has such a link. Across adoption levels, the scheme eliminates 92% of manoeuvre-induced false alarms and 96% of injected catalogue error. However, it cannot independently improve safety: a threshold rule calibrated for an uncertain catalogue responds insufficiently to an accurate one. A learned policy that observes the age of each neighbour’s data recovers this loss.

space situational awareness,Collision Avoidance,mega-constellations,multi-agent reinforcement learning,Graph Neural Networks (GNNs),conjunction screening