v3 weights (yolo11s, 204k merged plate dataset, val mAP50=0.944)

miner.py

@@ -450,8 +450,12 @@ class Miner:
         """Run the quad-4 tile pipeline and return RAW (pre-Soft-NMS)
         detections in original-image coordinates."""
         orig_h, orig_w = image_bgr.shape[:2]
-        OVERLAP_X = 35   # ~5% of 1408/2 (was 70/10%; halved to cut tile-seam duplicates)
-        OVERLAP_Y = 19   # ~5% of 768/2  (was 38/10%; halved to cut tile-seam duplicates)
+        # 2026-05-01: bumped to OVERLAP_X=55, OVERLAP_Y=32 after bench sweep on
+        # 33-task archive: archive HIGH R 0.8770→0.8934 (+1.64pp, +2 plates),
+        # phantom rate 11.7%→9.3% (-2.4pp). cid 62115 plate-1 IoU vs winner
+        # 0.520→0.849 (y-seam fix). starter R unchanged (saturated 0.9333).
+        OVERLAP_X = 55   # was 35; +1.64pp archive R from x-seam plate recovery
+        OVERLAP_Y = 32   # was 19; cures bbox-regression on plates spanning y-seam
         mx = orig_w // 2
         my = orig_h // 2