1
Why isolated scoring fails. A Transformer shares one residual stream, so units interact — two weak "bridge" units can be jointly indispensable.
2
Co-pruning curvature edges. One Fisher matrix — diagonal saliency, off-diagonal edges — recovered from just M single-unit ablations via a Gram product.
3
That it works. Strong across every capability class on Llama-3.1-8B at 20% and best on fragile code — the cross-module edges are provably decisive.