Layer-only ablation & interp — which neurons reduce EM, by method and layer
Scope: Layer-only — attribution + search restricted to only the trained mask layer's neurons. Each layer button = a mask trained on that layer. Pick algorithm → attribution method → layer to inspect the ablated neurons.
Method
k-sweep = ablate the top-K of a method's attribution ranking as a block. Greedy = search: add whichever neuron most reduces MCQ chose-misaligned. Scope Layer-only: attribution + search restricted to only the trained mask layer's neurons. Metric: MCQ single-token probe. Interp: max/min-activating on C4 + MCQ, C+/C−/grad×act.
Does ablation break the model? (a) share of the 2,400 free-form answers the judge calls coherent, before (black) and after ablating each method's greedy ceiling set (coloured) vs a size-matched random set (grey x); (b) that change per method, mean ± sd over layers; (c) per cell, how much EM was removed vs how much coherence was lost — points near the x-axis remove EM without breaking coherence; (d) the coding task: share of narrow-eval answers that are actual code (VULN or SAFE rather than NA) — a working coder stays flat. Judge: Claude Opus 5.
ALGORITHM
Attribution-method ranking (mean EM reduction across layers)
