Random images from the untouched 10,000-image official test set. Top line is the dataset truth; chips show each model's call. Inference runs in your browser right now (weights embedded, nothing leaves the page). Raise the abstain threshold and low-confidence calls go grey.
| Model | Accuracy | Macro F1 | NLL | Brier | ECE | Params |
|---|
NLL = negative log-likelihood, ECE = expected calibration error (15 bins). Lower is better for NLL, Brier, ECE. The augmented CNN trades a little clean accuracy for translation robustness — see the stress lab.
Bars: mean confidence per bin. Line: actual accuracy. A calibrated model hugs the diagonal.
Error rate on the kept predictions as the confidence threshold rises. Confidence helps — but it is not a safety net (see stress lab).
Real misclassifications from the untouched test set (augmented CNN). Shirt vs T-shirt/top vs Pullover confusions dominate — the weak classes in every model.
Apply a perturbation to a live batch of test images. Inference re-runs in your browser immediately. The full-audit numbers below were measured over all 10,000 test images during evaluation (seeded, reproducible).
| Perturbation | Model | Accuracy | Macro F1 | Mean conf (correct) | Mean conf (wrong) |
|---|
Key finding: confidence stays high even when the model is wrong under stress — abstaining on confidence alone does not catch shifted inputs.
Any PNG or JPG. The pipeline shows exactly what the models see: grayscale, centered, scaled to 28×28. White-on-dark is expected; tick invert if your sketch is dark-on-light.
The models know ten classes of 28×28 grayscale catalog clothing photos. A phone photo, a drawing, or a different garment style is out of distribution: expect confident nonsense. That is exactly the behaviour this workbench exists to expose.
Sorting-gate simulator, benchmark data only. The gate routes each garment to an auto-sort bin (its predicted class) when confidence clears the threshold, otherwise to manual review - the same way a confidence-gated conveyor line would act. This is a study in routing logic, not factory-ready QC: on 28×28 benchmark photos, confidence does not prove reliable physical sorting, defect detection, or calibrated safety. Under shifted inputs the models stay confidently wrong (see the stress lab), so a real gate misroutes exactly when it looks sure.
One image per row: 784 comma-separated pixel values 0–255, optionally followed by a true label 0–9 (header row allowed). Validated before anything runs.
Saved in your browser only (localStorage). Nothing is uploaded anywhere.
| When | Kind | Detail |
|---|