Deployed pipeline: SegGPT in-context drawer, K=8 query-retrieved supports (training-free) — test Dice 0.9667 (this cache: mean 0.9667 / median 0.9712, n=100).
▮ green = ground truth ▮ red = deployed prediction. Per-image Dice is burned on each tile; 6 fixed test images, chosen deterministically (spread across the dice-sorted test set).





