loading scene data…
The refusal direction, live
model
scene
induction
refusal
unlearning
emergence
SAE features
attribution graph
baseline
ablate dir
add dir
head that writes refusal (ring)
head that writes anti-refusal
REFUSED
answered
thermometer = refusal signal per layer
ablate: the signal drains into the well
drag = orbit · wheel = zoom · right-drag = pan
❚❚
0.5×
1×
2×
4×