Model Surgery: Blind Diagnosis, Autograft Repair from a Model's Own Checkpoint History, and the Economics of Where to Spend a Repair Budget

Localized damage to a trained transformer — a corrupted block of weights — is diagnosable, treatable, and its treatment is calculable in advance. This paper assembles the complete surgical loop at billion-parameter scale under sealed pre-registration. Diagnosis: a read-only per-cut deviation profile localizes single-block damage blind, refined through three pre-registered rules to a parameter-free first-nonzero tripwire that exploits the fact that a sick model is bit-identical to a healthy one upstream of the injury — nine-for-nine lifetime on pythia-1.4b, degrading gracefully to iterative peel on two simultaneous sites. Treatment: grafting the damaged block from the model's own 10,000-step-younger training checkpoint (an autograft, since the rollback checkpoint is the organ bank) and healing briefly beats plain healing three-for-three, with cross-entropy margins of 0.17–0.28 nats, recovering to about 1.004x healthy at every depth and inert at the wrong site. Economics, the paper's forward edge: two-seam surgery is safe; the recovery residual is heal-induced drift, not scar, yielding a stopping rule; a separable compatibility radius governs installation cost (a third of the model installs free from the adjacent checkpoint, cost climbing about 270x across a donor-distance ladder); and allocating a fixed repair budget to the damaged neighborhood beats whole-model healing by 2.4x at matched wall-clock. Instructive misses — a diagnosis rule that broke at scale, a sealed capacity-loss hypothesis the blind test overturned, a kernel-prices-surgery guess that failed — are reported at full price. All results are at 1.4B and below; 70B is named future work, not claimed. Nothing here concerns what a block means. Author contribution and use of AI: the research program and claims are the author's; experiments and drafting were done in collaboration with Claude, an AI system by Anthropic, under the author's direction, who verified the results and is responsible for the work. See the corresponding section in the PDF.