Multi-modal skin-lesion classification
Clinical and dermoscopic images combined with patient metadata, across 11 diagnoses including rare malignant classes.
Guardian recomputes a model's reported metrics from the evidence supplied with them. When the numbers and the evidence disagree, the claim is blocked and the discrepancy stays visible for human review.
–
Edit the declared value, or click a row to change its prediction. Guardian recomputes on every change.
sha256: –
Model reports carry numbers: accuracy, F1, AUC, error rates. Guardian checks each one against the files it was computed from, under the evaluation rules the team declared, and records the result in a signed evidence pack.
The team supplies the reported metrics, the prediction and label files behind them, and the evaluation rules: metric definition, threshold, rounding.
Guardian recomputes each metric from the rows. A difference is checked against the declared precision, so rounding never hides a real discrepancy.
Each claim is allowed, blocked as contradicted, or blocked for missing evidence. The verdicts, file hashes and rules go into a signed pack.
The declared value matches the value recomputed from the supplied rows at the declared precision.
The supplied rows give a different value. The claim cannot be used until the report or the evidence is corrected.
A file the claim depends on is missing or does not match the declared IDs. Guardian does not guess.
Our skin-lesion model on the ISIC MILK10k benchmark, written up with its limits next to its score.
| Claim | Value | Status |
|---|---|---|
| Benchmark score (macro Dice over 11 diagnoses)Public leaderboard, 479 test lesions | 0.625 | Verified · link |
| Overall rank, including entries with external dataLive leaderboard, November 2026 | Top 10 | Verified · link |
| Rank among entries without external data | 1st | Verified · link |
| Margin to the next two entries without external data0.611 and 0.609 | 0.014 – 0.016 | Within noise |
| Test-set noise, one modelOur resampling of out-of-fold predictions to the test size | ≈ 0.03 SD | Top 3 not separated |
| Training data and computeMILK10k training set only, public ImageNet weights | 1 × RTX 3090 | Declared |
Guardian grew out of our own modelling work, where rare classes, small test sets and leakage make reported numbers easy to overstate.
Clinical and dermoscopic images combined with patient metadata, across 11 diagnoses including rare malignant classes.
A public, reproducible package for JARVIS exfoliation energy: pinned commit, checksum-verified assets and a CPU replay script.
Point-in-time validation and cost-sensitive thresholds for rare-event detection, where a random split would leak the future.
Computational screening and free-energy calculations on a protein–protein interaction that is hard to drug.
Tell us which metrics you report and what evidence sits behind them. We will set up the private demo with a synthetic or anonymised example.
Write an email