From Saliency to Actionability: Layer Localization Is Protocol-Conditional

Layer-localization claims often compress distinct measurement and intervention questions into one layer number. We study a paired base-to-instruct contrast between Gemma-4-E4B and Gemma-4-E4B-it and ask when descriptive localization transfers to an internal actionability test. It does not transfer automatically. Eight measurement objects localize the same contrast to different depths, showing that magnitude, geometry, sensitivity, and intervention response are different estimands. Under the original raw-completion protocol, the target-gradient saliency peak at L9 yields no strict canonical first-token captures, while a late-layer patch at L30 captures 42/48 probes at alpha=0.5||h_L||. The magnitude is prompt-protocol dependent: matched diagnostics report 40/48 captures under raw-completion prompting but 6/48 under a chat-template diagnostic at the same late-layer setting. We therefore formulate localization as a protocol-conditional claim that must specify model context, prompt mode, measurement object, direction, intervention, scale, criterion, selection rule, uncertainty, and task context. The evidence supports actionability as a validation step for layer selection, not a universal layer law, formal mediation result, deployable steering method, or sequence-level correctness result.