Published photometric selections miss about one in five of 151 spectroscopically confirmed Little Red Dots: a learned selection and a candidate catalogue for nine JWST fields

Published Little Red Dot (LRD) samples inherit their selection rules' incompleteness. We test seven, reimplemented once on DAWN JWST Archive photometry of nine fields, against 151 spectroscopically confirmed LRDs: 149 from three published lists, 2 from a fixed spectral test. As re-implemented they recover 107 and miss 44, about three in ten. Of the 44, 11 are in a published LRD catalogue, missed only by our code. The other 33 are in no published photometric selection: 21.9 per cent of the 151, 18.7 per cent (23 of 123) at z>=4 and 21.1 per cent (31 of 147) at z> 3. We rank the catalogue on held-out sky with positive-unlabelled bagging of gradient-boosted trees anchored on 4,979 spectroscopic non-LRDs. At the union's burden it recovers 123 of them, including 20 of the 44 that no rule selects, against a shuffled-label floor reaching 14 on the 41 in support at the shuffled runs' own thresholds, and lets through 70 anchors against 46. Paired on the 147 in support, 20 against 3 (exact McNemar, p= 0.00049); it leads in all five regions at matched burdens. A label-free imitation of the rules recovers 116 there, against the union's 106 (p= 0.021) and the model's 124: ranking rather than cutting carries 10 of that 18-object gap. The labels' increment is not established. We release the model, code, tables and 1,207 unlisted candidates, none confirmed: a target list, not a discovery. In its equal-burden tier, 23.5 per cent of those observed have a secure z<=3. Deposited as a priority record while the manuscript is revised. The files (manuscript PDF, LaTeX source, and the code release as pushed to https://github.com/bakathefish/redress at commit 2f67aba) are restricted until then.