THE FEATURE ABLATION

Prove a component is load-bearing by removing it and watching the behavior break. Build a circuit whose part C literally implements a class split; knock it out (zero, then mean) and measure the collapse — while an unrelated part U knocks out to nothing. Necessity is not argued; it is demonstrated by absence.

sourceAblation / knockout in mechanistic interpretability — Morcos, Barrett, Rabinowitz & Botvinick, "On the importance of single directions for generalization" (2018), arXiv:1803.06959; single-unit & direction ablation. Circuit framing: Olah et al., "Zoom In" (Distill, 2020). Rendered, not quoted.

BLUE TEAM · BUILDS & DEFENDS
3
THE MODEL

A hand-built linear readout over two inputs x+, x- and two components. Every weight is fixed; every readout has a closed form.

y(x) = wC·aC(x) + wU·aU(x) + b
aC: [x+ → 2, x- → 0]  wC=2
aU: [x+ → 0.5, x- → 0.5]  wU=1  b=0

C is the key: it varies with the class, so it carries the split. U is constant across classes — pure offset, no class signal. Target metric = margin = y(x+)−y(x-). Intact margin = 4.

5
THE LINEAGE

The necessity test — the destructive twin of the-activation-patching. Patching writes a clean activation in and asks "is this enough?" (sufficiency). Ablation takes a component out and asks "does it break?" (necessity). Same circuit, opposite direction: one adds signal to find what suffices, one removes signal to find what is load-bearing.

7
THE WITNESS · live re-check

Re-runs the necessity property against the current engine state: mean-ablating C must drive margin to baseline 0. Trip the tamper in window 6 and this flips red.

witness idle

THE MACHINE
4
in ↓DATA IN

Two probes, one per class, and a choice of what to knock out and how.

inputs = { x+ , x- }  ·  components = { C=key , U=unrelated }
ablation ∈ { zero (→0) , mean (→E[a]) }

● ● ●
0
LITTHE PANEL · live knockout
rowy(x+)y(x-)marginbehavior
intact
ablated

● ● ●
8
out ↓DATA OUT

booting…

Proven at boot: knock out C (zero or mean) → margin 0, the split is GONE. Knock out U → margin stays 4, unchanged. Zero vs mean on C agree on margin (both 0) but leave different levels, differing by exactly wC·E[aC] = 2.

RED TEAM · ATTACKS & BREAKS
1
WALLTHE ADVERSARY

"A drop after ablation only proves the component participates, not that it is the mechanism. You broke the network's operating point — every downstream layer now sees off-distribution garbage. That is knock-on damage, not localization."

True on real nets: ablation is invasive. Off-distribution activations, compensating heads, and dead-neuron re-routing all muddy the read. This toy is linear and self-contained so the confound is absent here — that boundary is the honest claim, not a general one.

2
THE GRAVEYARD

"If ablating a component doesn't change the output, the component is useless."
→ It may be redundant, backed-up, or its effect may cancel at this operating point. Ablation shows this path is not solely necessary — not that the feature is unused.

"Mean-ablation is a gentler zero-ablation."
→ They answer different questions. Zero removes the whole activation; mean removes only the variation and keeps the average/bias. On C both kill the margin, but leave different output levels (Δ = 2).

"Bigger drop = more important."
→ Magnitude conflates the metric's scale with mechanism. Necessity is a structural claim (does the behavior survive?), read against a baseline — window 8. AMBER on trained nets the baseline itself is a modeling choice.

6
THE TAMPER · planted void

Swap the honest global mean for a leaky class-conditional "mean" that returns each class's own average — i.e. the original activations. The removal removes nothing; the behavior survives; C now looks unnecessary.

state: honest. The witness (7) confirms necessity holds.