Does Learned Feature Geometry Predict Detection Reliability?

Two ways of measuring geometry in a vision model have never been brought into one frame: the Zernike spectrum of a learned feature, and the localization error of a detector. This paper constructs the quantity that joins them. A detector already reports a displacement in units of the size of the thing that moved, the offset between a predicted box and its ground truth divided by the object's scale; we build the feature-side counterpart, the offset between an activation cluster and the same cluster recovered from an independently seeded retraining, divided by the cluster radius. Both are dimensionless and vary smoothly where IoU can flip on one pixel. The correspondence is constructed, not analogous: cross-seed clusters are matched by the harness's own assignment, cost and threshold, and one ResNet-50 trunk supplies both sides, trained as a classifier under eight seeds and then frozen inside a Faster R-CNN detector. The apparatus carries its own calibrations, and on RarePlanes a positive control separates one feature-side statistic from chance in the fixed-weight case, the cross-seed match rate, in 13 of 24 cells against 1.2 by chance; across retrainings it returns the chance value, in the shared canvas frame and after rigid registration, which lifts it fourfold for real pairs and for random positions alike. The displacement statistic the bridge rests on does not resolve the fixed-weight case, so its correlations cannot be read. Re-anchored to activation maps of real tiles, the dominant canvas pathologies are removed, measurement density rises by two orders of magnitude, and local feature density and spectral similarity show small associations with localization error, at $|{\rho }_S|$ near $0.1$, partly through object size and unresolved at the tile level: suggestive, not yet predictive. The detection side is unaffected: unfreezing the trunk lifts the decomposable reliability score from 0.862 to 0.925 on non-overlapping intervals.