A gated vision state-space model can erase closed tokens exactly, but its gate maps do not explain malignancy calls
Synopsis
In a lung-nodule malignancy concept bottleneck model, zeroing the mask on the selective state-space block's convolution input and giving closed tokens zero readout weight makes the concept and task logits exactly invariant to closed-token content; a path-by-path knockout over 15 cross-validated models reproduces this in every run, with the largest logit change exactly 0 when the conditions hold and between 4.1 and 6.7 when they do not. Gating is neither necessary nor sufficient, and the gate maps behave more like a mask-supervised support set than an explanation of the malignancy call.
Interpretation
A gated vision state-space model can carry a checkable exact-erasure guarantee: with the gate pattern fixed, the content of closed tokens cannot affect the concept or malignancy output. The selective step size is usually treated as the selection mechanism, on the assumption that a token the scan does not admit cannot influence the prediction. A path-by-path knockout over 15 cross-validated models on LIDC-IDRI gives a largest logit change of exactly 0 when the conditions hold and between 4.1 and 6.7 when they do not.
The guarantee comes from masking the convolution input, not from gating the step size. Removing the gate, the step-size gate or the residual gate one at a time leaves the guarantee intact, while removing the convolution-input mask breaks it. Across all 15 runs, removing the gate stays exact in 15 of 15, and removing the convolution-input mask is exact in 0 of 15.
The gated model matches an ungated Mamba concept bottleneck on task AUROC but has weaker concept quality. The gated model reaches a pooled out-of-fold AUROC of 0.8425 plus or minus 0.0085, against 0.8428 for the ungated Mamba bottleneck. One fixed patient-level five-fold split with three seeds, 15 models in total, with DeLong tests not significant in any of the three seeds.
Perspective
The guarantee holds for a fixed gate pattern and covers the concept logits and the task logit computed from them; in deployment the gate is itself a function of the image, so changing closed tokens can change the gate. All evidence comes from 764 nodules of LIDC-IDRI patients 0001–0600, one axial slice per nodule, under one fixed patient-level five-fold split with three seeds, with no external validation.
The guarantee is conditional: given this mask, the prediction cannot depend on what the mask hides, but it does not say that any content would produce the same mask, nor whether the open tokens are clinically meaningful. The gate maps fail a label-randomisation test, and concept-specific maps are only weakly reproducible even between two real-label seeds. The final model does not sit below the leakage level implied by true concept correlations, and the pre-registered F1 check fails.
