A Minimal-Pair Test Shows a Spectral Fork Diagnostic Tracks Prompt Surface Form

Prompt representation misalignment across layers is often taken as evidence for altered computation within a model. Here we examine one necessary condition for such claims: whether these misalignments track minimal interventions that change the answer, as opposed to surface altering (but not answer-altering) rewrites. To evaluate this we generate 100 grade school mathematics questions alongside five systematically controlled variants to each, including a single token answer altering edit, and multiple answer preserving rewrites. We find that in Gemma-4-12B-it, a spectral cluster-trajectory fork diagnostic tracks all formal answer-preserving rewrites (100/100) but none of the single-token answer-altering edits (0/100), which never elicited the original answer. The formal rewrite preserves the answer but drops accuracy from 70% to 56% ($p = .007$), while the answer-altering edit leaves accuracy intact. The controlled comparison holds surface change fixed: a proper-name substitution, identical to the answer-altering edit on every surface measure we compute, forks 4/70 while the answer-altering edit forks 0/100 (Fisher exact $p = .027$, one-sided). Across all 469 pairs, surface dissimilarity predicts forked status (McFadden $R^2 = 0.397$), and fork rate is monotone in token-count change across all five conditions. The answer-breaking coefficient is unidentifiable, since that arm has no fork events. While this study leaves open what aspect of surface altering edits is driving forked status, it suggests that when there's a simple answer altering control this is not a robust measure of answer relevant computation.