Criterion I5 — Confound Control

Validity type	Internal
Pass condition	The ablation effect is not explained by collateral disruption to non-circuit components
Evidence family	Causal
Minimum reporting	Component-specific ablation result; comparison to full-circuit ablation; mean vs. zero vs. resample comparison
Common failure mode	Running only mean ablation; not checking whether the effect is due to mean-field signal disruption

What this criterion requires

Confound control asks: is the observed effect due to the loss of the nominated component’s specific computation, or due to collateral disruption — unintended side effects affecting downstream components?

Satisfied when:

1. Component-specific ablation (ablating only the target, not adjacent components) produces the same degradation as the full-circuit ablation.
2. Mean ablation is compared to zero and resample ablation. Mean ablation changes the mean-field signal to all downstream heads — can cause cascade failures unrelated to the target component. If zero and resample produce comparable degradation, the mean-field confound is ruled out.
3. The effect is not due to hub disruption. Component-specific ablation paired with specificity test (I3) together rule out the hub confound.

The mean-field confound

When a component’s activation is replaced with the dataset mean, all downstream components receive the mean vector. For components with linear operations, the mean vector carries average training-distribution content, which may be task-relevant. This can suppress or amplify downstream effects unrelated to the ablated component’s specific computation.

Fix: Always run zero and resample alongside mean. If all three show comparable degradation, mean-field confound is ruled out. If mean ablation shows large degradation but zero shows small degradation — the effect may be mean-field disruption, not loss of specific computation.

Causal scrubbing as confound control

Causal scrubbing (c04causalscrubbing.py, A01 SCMPearl) replaces activations at each component with activations from a run where the causal variable is absent, while leaving all other activations unchanged. This is a component-specific counterfactual that controls for all pathway confounds simultaneously. A successful causal scrubbing result is the strongest single piece of confound-control evidence.

Minimum reporting rule

• Whether component-specific ablation was compared to full-circuit ablation.
• Whether ≥2 ablation methods were compared to check mean-field confound.
• Whether causal scrubbing was run (if so, report the result).
• If only one ablation method used: confound control is partial pass — flag explicitly.