loading scene data…
Attribution graphs: edges firing
model
scene
induction
refusal
unlearning
emergence
SAE features
attribution graph
excites (w > 0)
inhibits (w < 0)
feature node — size = influence
output logit (ignites last)
pulses run along an edge once the sweep passes its source layer
drag = orbit · wheel = zoom · right-drag = pan
❚❚
0.5×
1×
2×
4×