Criterion E1 — Intervention Reach

Validity type	External
Pass condition	A measurement confirms the target component’s activations changed in the predicted direction and magnitude
Evidence family	Causal
Minimum reporting	Activation delta at hook point (before vs. after intervention); direction confirmation
Common failure mode	Assuming the intervention reached the target because the code ran; never measuring the activation change

What this criterion requires

Intervention reach verifies that the intervention actually changed the target activations in the predicted way. Failure modes:

1. Wrong hook point: The hook name did not match the intended component.
2. Effect absorbed upstream: Skip connections or normalization layers partially mask the intervention.
3. Magnitude near zero: The component was near-inactive on test prompts; ablation delta is near zero regardless.
4. Wrong direction: A steering intervention added a vector not in the expected direction.

Satisfied when: activation value at the hook point is measured before and after intervention; delta is in the predicted direction; magnitude is non-trivial (not near zero).

Minimum reporting rule

• Measure and report activation value at the target hook point before and after intervention (sample of test prompts).
• Report mean absolute delta and confirm direction.
• If delta < 0.01 in normalized units: flag — the intervention may not have reached the target.

Why this is an external validity criterion

Intervention reach is a prerequisite for any external generalization claim. If the intervention did not reach the target reliably in the original setting, there is no basis for cross-model or cross-task generalization. It also functions as a measurement validity check on the experimental procedure.