{"v3":{"calibration":{"drawings":192,"top1":0.75,"top5":0.9583333333333334,"meanReciprocalRank":0.8390450261544012},"fresh":{"drawings":192,"top1":0.765625,"top5":0.9583333333333334,"meanReciprocalRank":0.8451594268000521},"mixed":{"boards":96,"targets":192,"top5":0.5677083333333334,"meanReciprocalRank":0.3637819073547357,"targetVsAbsentAUC":0.7627840909090909}},"v4":{"calibration":{"drawings":192,"top1":0.8020833333333334,"top5":0.953125,"meanReciprocalRank":0.8700644841269843},"fresh":{"drawings":192,"top1":0.8020833333333334,"top5":0.9583333333333334,"meanReciprocalRank":0.8687479332010583},"mixed":{"boards":96,"targets":192,"top5":0.6041666666666666,"meanReciprocalRank":0.37172385437006694,"targetVsAbsentAUC":0.7954545454545454}},"dinov2":{"checkpoint":"onnx-community/dinov2-small","revision":"8b1f705a3a7f6f062f6bdd21986c1583d3ef105d","sha256":"c179f8f7f592449c4c1bca4cd124a7538021428c5ffb89afde9503935b197efb","backend":"1.31.1","provider":"CPUExecutionProvider","threads":2,"batch":4,"samples":672,"inferenceSeconds":46.666,"method":"L2 of final CLS token; coordinate median of 12 support vectors per class, L2 again, cosine ranking. Equal input scaling; no augmentation.","labels":["canoe","eye","dishwasher","waterslide","mountain","dresser","scissors","garden hose","rabbit","rain","lightning","circle","broom","pickup truck","horse","ceiling fan","cello","basket","stop sign","carrot","smiley face","strawberry","cat","tornado"],"splits":{"calibration":{"count":192,"top1":0.791667,"top5":0.953125,"mrr":0.867188,"perClass":{"canoe":{"top1":7,"test":8},"eye":{"top1":6,"test":8},"dishwasher":{"top1":7,"test":8},"waterslide":{"top1":7,"test":8},"mountain":{"top1":5,"test":8},"dresser":{"top1":6,"test":8},"scissors":{"top1":8,"test":8},"garden hose":{"top1":2,"test":8},"rabbit":{"top1":6,"test":8},"rain":{"top1":5,"test":8},"lightning":{"top1":7,"test":8},"circle":{"top1":8,"test":8},"broom":{"top1":6,"test":8},"pickup truck":{"top1":8,"test":8},"horse":{"top1":4,"test":8},"ceiling fan":{"top1":5,"test":8},"cello":{"top1":7,"test":8},"basket":{"top1":5,"test":8},"stop sign":{"top1":8,"test":8},"carrot":{"top1":6,"test":8},"smiley face":{"top1":8,"test":8},"strawberry":{"top1":8,"test":8},"cat":{"top1":6,"test":8},"tornado":{"top1":7,"test":8}}},"fresh":{"count":192,"top1":0.802083,"top5":0.973958,"mrr":0.876218,"perClass":{"canoe":{"top1":7,"test":8},"eye":{"top1":6,"test":8},"dishwasher":{"top1":6,"test":8},"waterslide":{"top1":8,"test":8},"mountain":{"top1":5,"test":8},"dresser":{"top1":8,"test":8},"scissors":{"top1":3,"test":8},"garden hose":{"top1":3,"test":8},"rabbit":{"top1":7,"test":8},"rain":{"top1":6,"test":8},"lightning":{"top1":7,"test":8},"circle":{"top1":6,"test":8},"broom":{"top1":7,"test":8},"pickup truck":{"top1":8,"test":8},"horse":{"top1":7,"test":8},"ceiling fan":{"top1":6,"test":8},"cello":{"top1":6,"test":8},"basket":{"top1":7,"test":8},"stop sign":{"top1":8,"test":8},"carrot":{"top1":4,"test":8},"smiley face":{"top1":8,"test":8},"strawberry":{"top1":7,"test":8},"cat":{"top1":7,"test":8},"tornado":{"top1":7,"test":8}}}},"mixed":{"boards":96,"targets":192,"negativeComparisons":4224,"mrr":0.386018,"top5":0.59375,"targetVsAbsentAUC":0.798769,"seconds":6.205}},"timing":{"v3":{"warmup":64,"samples":736,"medianMs":0.11161799999990762,"p95Ms":0.15942100000006576,"runtime":"v24.22.1","environment":"Node server CPU, single process, 10-stroke boards, includes both prompt scores; excludes network, D1 and startup."},"v4":{"warmup":64,"samples":736,"medianMs":1.2688169999996717,"p95Ms":1.768670999999813,"runtime":"v24.22.1","environment":"Node server CPU, single process, 10-stroke boards, includes both prompt scores; excludes network, D1 and startup."},"dinov2":{"backend":"ONNXRuntime 1.31.1 CPUExecutionProvider, two threads, batch one","warmup":5,"samples":32,"medianMs":82.33,"p95Ms":107.59,"meanMs":83.49,"environment":"Shared executor. Rasterization, normalization and CLS inference on ten-stroke boards. Timing is local, not deployed latency."}},"protocol":"Matched 24 deterministically selected QuickDraw words. Each model used 12 support sketches per word to refit a coordinate median of L2-normalized vectors. Two domain models additionally learned their image encoders on WeDraw training originals; DINOv2-small was frozen, without fine-tuning. For each label 8 independent post-v3 QuickDraw records were compared to all 24 medians, then 96 synthetic 5+5 boards were built from these same records. Closed 24-way retrieval; this is not the 413-way game and measures no human judgments."}