One row per named move type. Each model is an attention-MIL net — expand a row to see its actual decision: which frames it attends to (the trimmed move), and which joints drive it. Train with npx tsx scripts/train-move-classifiers.ts --mil. Scan for poison labels with npx tsx scripts/train-move-classifiers.ts --label-quality — suspects show a red banner below. The ⚠ button re-runs the cheap checks (contradictions, duplicates, source sync) live without the k-fold scan.
Every batch is saved on the server. Open any scan to review its ranked matches later.
Choose a positive example below to start a saved scan.