Frame Dependence: Circuit Robustness Under Task-Preserving Distribution Shift

Measures whether a circuit discovered on a diverse set of sentence frames remains faithful on frames it was not discovered on, and finds the question cannot be answered at the resolution standard top-k discovery provides. Across four task families on Llama-3.2-3B and Pythia-1.4B, transfer penalties are near zero, but planting a known perturbation shows that swapping one head of ten moves faithfulness by 0.03-0.51 - more than any observed penalty. The minimum detectable effect is a two-to-five head circuit difference. Seven of eleven penalties are negative, a property of the selection procedure rather than of frames. Faithfulness rises monotonically with circuit size, reaching 1.198 at twenty heads where ten give 0.390, so a previously published 0.663 to 0.228 collapse records a circuit-size artifact rather than frame dependence.