Formal current Test report · 2026-09-07

Segmentation across four dev domains

Sealed current-Test scores from the frozen Support-fitted Harness routes, with three fixed-weight NN references for ISIC, Kvasir, CXR, and BUSI. The twelve comparison panels use the same current Test cases for Harness and fixed NN references.

224 × 224 evaluationSupport budgets by image countthree NN familiesno Test adaptation

Current and fixed-weight NN Test scores by Support budget

N is the exact number of labeled Support images. All displayed domains keep Support/Test images and groups separate. CXR uses the common overlap-excluded cohort described below. Group-balanced Dice first averages Test images within each group. Harness scores use the sealed Test reports linked below; U-Net and SegFormer use receipted three-seed ensembles; nnU-Net uses its five fixed fold checkpoints. Full precision is retained in the current source JSON and the fixed NN source JSON, with nnU-Net scores and receipts.

ISIC 2018
Method	N=10	N=20	N=30	N=40	N=50	Test images / groups
Harness · website	0.866	0.864	0.865	0.868	0.871	100 / 100
Fixed NN · U-Net	0.780	0.817	0.793	0.814	0.826	100 / 100
Fixed NN · SegFormer	0.777	0.836	0.845	0.845	0.859	100 / 100
Fixed NN · nnU-Net	0.750	0.836	0.835	0.832	0.846	100 / 100

Kvasir-SEG
Method	N=10	N=20	N=30	N=40	N=50	Test images / groups
Harness · selected decoder Test	0.779	0.799	0.822	0.810	0.817	100 / 100
Fixed NN · U-Net	0.625	0.710	0.791	0.820	0.826	100 / 100
Fixed NN · SegFormer	0.694	0.721	0.831	0.851	0.861	100 / 100
Fixed NN · nnU-Net	0.598	0.647	0.683	0.728	0.757	100 / 100

CXR
Method	N=10	N=20	N=30	N=40	N=50	Test images / groups
Harness · Support-selected	0.937	0.948	0.954	0.955	0.958	106 / 106
Fixed NN · U-Net	0.953	0.967	0.968	0.971	0.973	106 / 106
Fixed NN · SegFormer	0.957	0.966	0.968	0.969	0.971	106 / 106
Fixed NN · nnU-Net	0.958	0.966	0.969	0.970	0.973	106 / 106

CXR uses the same 106 Test images at every budget and for every method. We excluded all 12 images whose ID or group appears in any of the five Support budgets. NN scores were recomputed from the original per-image results without retraining. Excluded IDs and cohort · Recomputed scores.

BUSI
Method	N=11	N=23	N=38	N=49	N=60	Test images / groups
Harness · Support-selected	0.611	0.573	0.573	0.597	0.603	125 / 100
Fixed NN · U-Net	0.516	0.673	0.716	0.772	0.799	125 / 100
Fixed NN · SegFormer	0.643	0.709	0.753	0.787	0.807	125 / 100
Fixed NN · nnU-Net	0.453	0.565	0.622	0.715	0.722	125 / 100

BUSI uses the Support-selected joint image-presence then patch-Ridge method on 11/23/38/49/60 labeled images. The original archive labels these budgets by 10/20/30/40/50 groups; N above denotes images. See the audited sealed Test report for the complete score, prediction-seal, and Support-selection bindings; the gallery case bindings are in the linked manifest below.

CXR Harness scores use the sealed 106-image Test report. All five budgets use the Support-selected frozen dense label-transfer method; the independently audited N10/N20/N30/N40 improvements are newly verified, while the independently audited N50 result and presentation gallery are reused unchanged. All five CXR axes remain below the matched fixed-NN target and strict 0.010 tolerance. BUSI Harness scores use its audited sealed Test report for the Support-selected joint image-presence then patch-Ridge method. Green marks the best available score per budget. The preserved four-row source artifact contains earlier CXR results and is not used in these tables.

Green score text marks the best full-precision value across Harness, U-Net, SegFormer, and nnU-Net at each budget; displayed values are rounded to three decimals. All three NN families use the same Support cohort and current Test cases at each N. The nnU-Net comparison uses the official five-fold final-checkpoint ensemble; its training-fold validation scores are not reported as Test scores. Fixed U-Net/SegFormer score sources and weight digests are recorded in nn_baseline_scores.json. The fixed archive receipt is adc91bca…; the aggregate receipt is dfb9b5dd….

Current Test comparison gallery

Each four-panel composite uses 50 Support images for ISIC/Kvasir/CXR or 60 for BUSI, with cases selected by quartiles of current Harness Dice after sorting by (Dice, id). The Harness input uses the ISIC native image with LANCZOS or the prepared 224 × 224 view for the other domains; the fixed NN uses the original 224 × 224 BILINEAR terminal view and float32 three-seed probabilities thresholded at 0.5.

ground truthcurrent Harnessfixed NN ensemble

The ISIC/Kvasir PNGs and source bindings are recorded in the current gallery manifest. The CXR and BUSI galleries below use their separate matched Test reports.

ISIC 2018

current Harness + fixed N=50 SegFormer

Current ISIC lower-quartile case: input, ground truth, Harness, and fixed NN ensemble — Lower quartile · ISIC_0016344 · Harness Dice 0.844 · fixed NN SegFormer Dice 0.949

Current ISIC median case: input, ground truth, Harness, and fixed NN ensemble — Median · ISIC_0036215 · Harness Dice 0.921 · fixed NN SegFormer Dice 0.817

Current ISIC upper-quartile case: input, ground truth, Harness, and fixed NN ensemble — Upper quartile · ISIC_0019030 · Harness Dice 0.963 · fixed NN SegFormer Dice 0.819

Kvasir-SEG

current Harness + fixed N=50 SegFormer

Current Kvasir lower-quartile case: input, ground truth, Harness, and fixed NN ensemble — Lower Quartile · cju1cu1u2474n0878tt7v4tdr · Harness Dice 0.708 · fixed NN SegFormer Dice 0.753

Current Kvasir median case: input, ground truth, Harness, and fixed NN ensemble — Median · cju5ufn3skquf0818dhapnhba · Harness Dice 0.946 · fixed NN SegFormer Dice 0.911

Current Kvasir upper-quartile case: input, ground truth, Harness, and fixed NN ensemble — Upper Quartile · cju5von04litr08718j8po40a · Harness Dice 0.970 · fixed NN SegFormer Dice 0.932

Chest X-ray

50 Support images · current Harness + fixed U-Net

CXR lower quartile case: input, ground truth, Harness, and fixed NN ensemble — Lower quartile · MCUCXR_0061_0 · Harness Dice 0.955 · fixed NN U-Net Dice 0.979

CXR median case: input, ground truth, Harness, and fixed NN ensemble — Median · MCUCXR_0028_0 · Harness Dice 0.965 · fixed NN U-Net Dice 0.979

CXR upper quartile case: input, ground truth, Harness, and fixed NN ensemble — Upper quartile · MCUCXR_0094_0 · Harness Dice 0.972 · fixed NN U-Net Dice 0.974

CXR panels are selected from the common Test cohort of 106 images, with no Support ID/group overlap. CXR gallery sources.

BUSI

60 Support images · current Harness + fixed SegFormer

BUSI median case: input, ground truth, Harness, and fixed NN ensemble — Median · Harness Dice 0.752 · fixed NN SegFormer Dice 0.872

BUSI upper quartile case: input, ground truth, Harness, and fixed NN ensemble — Upper quartile · Harness Dice 0.917 · fixed NN SegFormer Dice 0.944

BUSI panels use the same current Test cases for both methods, selected at the lower quartile, median, and upper quartile of Harness Dice. BUSI gallery sources.

Recipe and provenance

ISIC Harness

gpt-5.6-sol structured polygon draws followed by the preregistered Support-fit numpy genome. Harness gallery input is the native ISIC image resized with LANCZOS.

Kvasir Harness

The independently verified selected decoder is scored on the current 224 × 224 terminal view. HQ is retained at N=10/20/40 and the authenticated SAM2 decoder at N=30/50; the score, prediction, and audit bindings are recorded in the gallery manifest.

Fixed NN

U-Net and SegFormer each use three frozen seed checkpoints per N. nnU-Net ensembles five frozen fold checkpoints per N. All scores use current Test cases, without retraining. Each gallery uses the strongest displayed NN family at the largest Support budget: SegFormer for ISIC/Kvasir/BUSI and U-Net for CXR.

The fixed baseline archive is descriptive Test evidence. It is not an adaptation signal. CXR and BUSI use their completed formal Test score receipts.