Three fine-tunes of the same public, CC0-licensed CoAtNet-384 checkpoint, each trained on a differently-preprocessed corpus, rank-blended together score 0.9288 on the 58 gold-labeled studies - up from 0.9255 for the strongest one alone. Getting the base recipe right took fixing five real bugs in an earlier from-scratch replication; getting the ensemble right took finding the exact per-checkpoint preprocessing config each snapshot was actually trained on, not assuming they all share one recipe.
\ndreaddevelopment/knee-mri-twelve-findings-from-a-single-model, a 0.924-public-LB notebook) turned up five concrete bugs: wrong normalization (buffers guessed inside the model instead of the real external ImageNet mean/std), a wrong attention-head activation and an unnecessary presence-masking path, missing DICOM modality LUT and MONOCHROME1 handling, percentile normalization computed per slice instead of per slot, and a 15-window sample instead of the real 42 drawn from a 64-slice volume. Fixing all five and reloading with strict=True brought the gold-only score to 0.9213, then a window-count sweep found a 56-60 plateau above the shipped 42, landing on 0.9255.widedense, maxspan - the one used above, widefov, fullspan, native384, finespacing, native384dense). An earlier attempt at blending four of these (\"arms\") failed - solo AUCs came back near random - because every arm was fed through the maxspan recipe regardless of what it was actually trained on. Reading a competitor's public notebook that blends these same checkpoints turned up their exact per-checkpoint build resolution, slot layout and depth-span settings. Rebuilding two arms (native384 and native384dense) with their own correct settings brought both to genuinely strong solo scores (0.9116 and 0.9170), and a rank-blend of all three reaches 0.9288.strict=True on each checkpoint load only succeeds on an exact architecture match; it isn't a formality, it's the actual proof each checkpoint fits this class.maxspan checkpoint wants a 64-slice volume across five slots (18/14/12/8/12) built at 336px; native384 wants a smaller 44-slice volume (12/10/8/6/8) built natively at 384px; native384dense keeps the 64-slice layout but also builds natively at 384px. Every slice is still cropped to a 140 mm box from the DICOM pixel spacing, modality-LUT corrected, and MONOCHROME1-inverted where needed, and every slot's slices share one percentile normalization rather than being rescaled independently.dreaddevelopment/raptor-knee-maxspan, dreaddevelopment/raptor-knee-native384dense, and dreaddevelopment/raptor-knee-native384 (all CC0-1.0).maxspan settings come from the checkpoint author's own published inference notebook; the native384 and native384dense settings come from a competitor's public notebook that blends the same three checkpoints, read directly rather than reverse-engineered. An earlier attempt to blend four other arms without their correct per-checkpoint settings produced near-random solo predictions - the lesson held here too, just with the right settings in hand this time.maxspan-v5 0.9255 (the same checkpoint the previous version of this notebook shipped, scoring 0.928 on the real leaderboard at its 42-window setting), native384dense-v10 0.9170, native384-v8 0.9116. Rank-blended at the weights found above: 0.9288. This is the first ensembling gain of any real size found in this project since the base recipe was fixed - every other attempt at adding a second model added at most a few thousandths of an AUC point, because none of those secondary models were anywhere near raptor's own accuracy. These three checkpoints all share raptor's own architecture and general training recipe, just fit on differently-preprocessed corpora - close enough in strength that combining them adds real information instead of diluting it.widefov v6 0.9108, fullspan v7 0.9063, finespacing v9 0.8985 solo - all decent, all driven to zero weight by the joint search); the competitor's own real out-of-fold predictions for a 5-fold DINOv2 model (0.8400 solo, +0.0000 blended) and three RadImageNet-ResNet50 variants (0.83-0.86 solo, +0.0006 blended, best case); and a properly task-fine-tuned OrthoFoundation-L (a DINOv3 model pretrained on ~1.25M knee images), taken from a 0.75 frozen-embedding probe to a real 0.8124 solo AUC via three epochs of attention-pooling fine-tuning - still only +0.0004 blended. Four checkpoints in the raptor family remain untested (widedense v4/v4-swa and two others) for lack of a verified per-checkpoint recipe - not attempted, to avoid repeating the near-random-prediction failure that a wrong config produces.