Single-opposition objects (3–5 nights) whose observations a machine-learning model reads as
likely holding more than one body: tracklets of different asteroids linked into one
designation. Each score is at or above the model's operating point (0.93),
where no control arc in its external test set was flagged while about half of the known
mislinkages were. Scores read as 0.97+ very likely, 0.93+ likely (this list), 0.70+ possible,
0.27+ small chance (higher than 99% of real arcs), 0.02+ very remote chance, below that no sign. A score is a ranking, not a probability. Click an object to see why
the model flagged it: each feature's contribution, adding up to the score.
The list is fixed in the code (not generated live); the “now” column is the same model run on
the published observations when the list was exported, so a fixed or re-observed object shows
there.
The model is a gradient-boosted tree classifier over 26 features of the arc: how its nights
and gaps are built, the photometry of each tracklet, one orbit fit to every night, a refit
that leaves out the smaller side of the largest gap and predicts it, and a night-by-night
chain grown from the best-observed pair of nights. Explanations are exact TreeSHAP values
in log-odds. It was trained on main-belt to Hilda arcs; objects with q < 1.3 au or
a > 5.5 au are not scored here. “Reviewed chimera” and “known mislinkage” mark objects
already judged by hand; their scores come from a model that never saw that judgement.