Efficient Mechanistic Circuit Analysis of Open-Weight LLMs on Apple Silicon

Version 2 (correction). The fall in faithfulness from 0.663 to 0.228 (Llama) and 0.854 to −0.028 (Pythia) on the 15-template dataset, previously framed as single-template fragility, is an artifact of holding the circuit at ten heads — a size inherited from the single-template setting. Sweeping circuit size on the same diverse data reaches F = 0.715 at k = 15 and 1.193 at k = 20: a sparse circuit for diverse IOI exists, and ten heads is too few to express it. No measurement is withdrawn; the abstract, contribution 5, the template section and the conclusion now state the size effect and carry the sweep. What the comparison establishes is that circuit size must be swept and reported.Replication and extension of the GPT-2 Small IOI circuit analysis on Llama-3.2-3B and Pythia-1.4B using an MLX activation-patching harness on a single Apple Silicon machine. Circuits are faithful on the standard single-template dataset (F=0.663 and 0.854) but collapse on a 15-template dataset (0.228 and -0.028) while both models continue to perform the task above 90% of their original logit difference. Single-template IOI circuits are properties of a sentence frame rather than of the task.