1
Why one-shot collapses. A single second-order map, drawn at the dense model and extrapolated across the whole budget, breaks down super-exponentially at high sparsity.
2
Anchored continuation. Separate a fixed dense anchor from a moving expansion point; re-estimate curvature along the path in K steps — errors add, not multiply.
3
That it works, inspectably. Up to 16.8× (and ~270×) lower perplexity than one-shot at 50% sparsity — changing only the mask. One-shot is exactly K=1.