Logical anomaly detection stops rewriting only the question: CASCADE++ scaffolds the image per constraint and beats full-shot training with three references

The bottleneck in logical anomaly detection sits on the perception side: making the check question more elaborate and letting the model reason longer does not recover the verdict, while rewriting the image the model sees for each constraint does. Earlier checklist methods decomposed only the text, generating an independent question per constraint while every check read the same unmodified query image. In a training-free comparison with backbone and question fixed, chain-of-thought and self-refinement textual guides improved on the text-only baseline by no more than a point on binding and calibration constraints alike, whereas a single disambiguation overlay raised balanced accuracy on the binding Count constraint to a level no textual guide approached.

CASCADE routes by constraint type: binding constraints need visually similar instances separated, calibration constraints need a ruler consistent across images, and each type gets its own overlay. Earlier visual scaffolds took their parameters from image space, numbered instances by image position and fixed grid spacing, independent of the constraint's logical structure and the normal range defined by few-shot references. A synthetic probe shows raw-image accuracy counting identical circles collapsing past a certain count while per-circle coloring keeps it stable across three backbones; readings of a fixed-length bar spread noticeably wider across separate renderings than within one repeated image, and an object-anchored two-dimensional grid holds accuracy as the violation margin shrinks while single-axis and image-frame overlays fail.

CASCADE++ replaces the fixed scaffold defaults with configurations derived from the constraint structure and the reference distribution, flipping the image-level result from below the baselines to ahead of them. Base CASCADE left the scaffold configuration at a fixed default and fell below both reproduced baselines on MVTec-LOCO. On the GPT-5-image-mini backbone with three references, CASCADE++ reaches a higher AUROC than the strongest reproduced baseline under the same backbone and reference budget (LogSAD†) and than the full-shot LogSAD, with a small standard deviation across three random reference sets; the same scaffold transfers to Gemini-3.1-Flash-Lite, Qwen3-VL-32B and Qwen3-VL-8B.

The demonstrated scope is five product categories of MVTec-LOCO and ten manufacturing scenarios of VID-AD, with all constraints falling into the binding and calibration types and each type matched to one scaffold; the framework targets near-line inspection rather than in-line control, since verifying one image takes far longer than the millisecond latency of feature-based detectors.

The number of references governs how reliably the calibration grid is recovered: a single reference cannot span the normal range, a too-narrow grid flags normal queries as violations, and the estimate stabilizes at three references with negligible further gain. Segmentation quality matters asymmetrically: boundary dilation within 10 and 20 pixels stays within noise, while splitting or merging instances clearly lowers average accuracy. Constraints that combine both perceptual operations are not yet covered.