Get Needs Labeling

Active-learning queue — verdicts where calibrated p_ship sits in the uncertainty band [p_low, p_high]. These are the highest-ROI examples for human labeling: they're the ones the model is least sure about, so each label corrects the calibration the most.

Returns up to limit rows, oldest-first (we label them in the order they were emitted to keep the holdout time-balanced).

Auth: paid+ only. Anon callers 401.