Skip to content
Validity typeConstruct
Pass conditionThe proposed circuit does not score highly on unrelated tasks under the same instrument and intervention strength
Evidence familyRepresentational (cross-task IIA), Behavioral (cross-task faithfulness)
Minimum reportingIIA or faithfulness on ≥1 unrelated control task; ratio of on-task to off-task score
Common failure modeReporting only the target-task score; never testing whether the same circuit scores well on completely unrelated behaviors

Task specificity is the discriminant side of construct validity. It asks whether the circuit is specific to the claimed computation, or is a general-purpose structure that would score well on any task.

Satisfied when:

  1. The circuit is tested on ≥1 unrelated control task. Different in computational structure, not just a variant. For SVA: Greater-Than is a good control. For IOI: Greater-Than or gendered pronoun.
  2. The circuit’s score on the control task is substantially lower. On-task:off-task ratio ≥ 2:1, or off-task score not significantly above chance using the same instrument and strength.
  3. Cross-task IIA contamination is near zero. Using the same subspace alignment, the IIA on the control task should be near baseline.
R = IIA(target task) / IIA(control task)

R = 1.0 means the circuit scores equally on both — task specificity fails. R ≥ 2.0 with off-task score not significantly above baseline is a reasonable pass threshold.

The SVA circuit claim at L8.MLP requires: IIA on Greater-Than or IOI (same DAS-IIA setup). If L8.MLP shows IIA = 0.48 on SVA and IIA = 0.10 on Greater-Than (R = 4.8), task specificity is satisfied. If IIA = 0.45 on Greater-Than (R = 1.07), the construct is “general information-processing locus,” not SVA-specific — a weaker and different claim.

  • Control task, intervention strength used for control.
  • Off-task score with same precision as on-task score.
  • Specificity ratio.
  • If off-task score is unexpectedly high, discuss — may mean the construct is under-specified.