Both reruns use one bio; each keeps different nodes fixed.
Unedited run on x₀: a = 1.00, MLP = 1.30, backup = 0.00, y = 1.02 → nurse
path patching: only a's path to y zeroed, grey nodes held at their unedited-run values
x₀ one bio · a one head's output on it · ã a set to 0 · backup a head that stays idle while a is active and fills in when a drops · y score for nurse, above 0 reads nurse, otherwise surgeon · numbers in the boxes are each node's value in that run.
Toy network: MLP = max(0, 1.2a + 0.1), backup = strength × max(0, 1 − a), y = 1.5a + 0.4 MLP + 1.2 backup − 1.
Unedited a: gender 1.00, cue 1.00, y = 1.02 → nurse
a here the bio's activation vector at one layer · gender, another cue two directions in a that travel together in the bios; hidden: the analyst sees a and the probe, not these axes · 0° the probe points exactly along gender.
probe = 0 the line where a has no probe component; the cue axis is where gender = 0 · dashed what each removal takes out of a.
Toy readout: y = 0.4 × gender + 1.2 × cue − 0.58, above 0 reads nurse; shaded where it reads surgeon.