Figure — the ablation lattice, 512 measured coalitions
Switch heads off and watch the circuit degrade
Every one of the 29 = 512 combinations below was measured on a T4: keep the selected heads
intact, mean-ablate all 144 others, and record the mean logit difference. Nothing here is interpolated —
click any set and you get the number that run produced.
corrupt floor
clean baseline
mean logit difference
The 9 heads are the corrected circuit, found by greedy backward elimination over all 144 heads
rather than by a threshold. Shapley values come from this same lattice.