What we tried.
What survived.

The early LoRA comparison tested a prototype-based objective against feature consistency. The prototype objective gave the stronger result, motivating the later support-set prototype formulation rather than treating teacher/student consistency as the default adaptation target.

• Prototype: AUROC 85.69%, AP 82.63%, F1 81.28%.
• Consistency: AUROC 81.98%, AP 76.59%, F1 78.57%.
• This is a VisA experiment, not directly comparable with the later MVTec AD 2 table.

The proxy-validation line uses four folds for a four-shot support set. In each fold, three normal support images train the adapter and the held-out normal image validates it. This provides a repeatable way to inspect normal-feature behavior while keeping synthetic defects disabled for the reported MVTec AD 2 comparison.

• Four leave-one-normal-shot-out folds: 3 support images for training, 1 held out for validation.
• The validation signal is diagnostic; it is not proof that anomaly ranking improves.
• The single-adapter route trains a final normal-only proxy adapter after fold validation.

On the MVTec AD 2 seed-0 comparison, this did not beat the frozen baseline on AUROC or AP, though it was slightly faster and raised F1 by 0.75 percentage points. This branch became the common foundation for the later proxy experiments, not a final performance win.

The frozen DINOv2 feature bank is the control for the subsequent MVTec AD 2 experiments. Across the eight public categories, it outperformed the single proxy adapter on AUROC and AP. This is an important constraint: improving normal compactness or fitting a proxy objective does not guarantee better separation between normal and anomalous queries.

All values in this section are seed-0 macro results. They are a useful development reference, not a multi-seed statistical claim.

The corrected implementation freezes each fold's final-epoch adapter as a teacher. A new student is then trained for a full 40 epochs. For every support image, one student view is matched independently to four independently augmented teacher views; the four patchwise cosine losses are averaged rather than collapsing the teachers into one averaged target. Best-validation epochs are diagnostic only, not the teacher checkpoints.

Measured result · MVTec AD 2

• Versus frozen baseline: AUROC −4.24 pp, AP −3.40 pp, F1 +0.34 pp.
• Versus the single proxy adapter: AUROC −1.42 pp, AP −0.16 pp, F1 −0.41 pp.
• Versus the earlier averaged-target consensus run: AUROC −0.71 pp and AP −1.10 pp, with F1 +0.22 pp.
• Fabric AUROC fell from 56.13% in the legacy consensus result to 45.19% in the corrected result; fruit_jelly improved from 72.08% to 79.50%.

Why this became a dead end

The student objective rewards teacher imitation plus normal-bank compactness. With synthetic defects disabled, there is no explicit term pushing anomalous queries away from the normal region. A falling teacher-matching loss therefore demonstrates imitation, not improved anomaly separation. The fabric regression reinforces that the cross-fold compromise can remove useful discriminative variation.

Decision: do not increase consensus weight or training duration based on this run. Teacher disagreement and student-to-each-teacher losses are useful diagnostics before any further consensus experiment. Results are one seed only.

Instead of forcing four fold adapters into one student feature space, this branch retains the four fold-best models selected by proxy validation. Each model independently encodes all four normal shots to build its own feature bank, scores the query in that model's geometry, and contributes a patchwise anomaly map. The maps are averaged. Embeddings from different adapters are not concatenated.

Measured result · MVTec AD 2

• Versus frozen baseline: AUROC +0.63 pp, AP −0.43 pp, F1 +1.76 pp.
• Versus corrected consensus: AUROC +4.87 pp, AP +2.97 pp, F1 +1.42 pp.
• Inference is about 3.35× slower than the frozen baseline and 3.15× slower than corrected consensus.
• Per-adapter diagnostic JSON/CSV files were generated in local result folders, but were not committed with the run logs; image-level disagreement and false-positive/false-negative analysis still requires those files.
• Category AUROC: can 29.00%, fabric 63.37%, fruit_jelly 80.58%, rice 79.76%, sheet_metal 73.24%, vial 76.03%, wallplugs 33.63%, walnuts 85.11%.

Why it survived

It preserves each fold adapter's own feature geometry rather than distilling potentially disagreeing representations into a single student. It is the best observed AUROC and F1 in the current MVTec AD 2 comparison, but it is not a clean win on every metric: frozen features still have slightly better AP, and the ensemble is substantially slower.

Decision: keep as the lead candidate, inspect the per-adapter diagnostics and false-positive/false-negative lists, then repeat paired seeds 1 and 2 (preferably more) before making a general performance claim. Alternative score-fusion variants remain exploratory because choosing among them on the same public test labels would introduce test-set selection bias.