Skip to content
Validity typeConstruct
Pass conditionMultiple independent instruments from different evidence families nominate the same components
Evidence familyMeasurement (cross-instrument agreement)
Minimum reportingJaccard similarity between instrument outputs; both instruments’ component lists; interpretation of any disagreement
Common failure modeRunning multiple instruments but not checking whether they agree; treating agreement as assumed

Convergent validity is the core of this entire project. Running one instrument and claiming a circuit is not enough.

Satisfied when:

  1. ≥2 instruments from different evidence families have been run. Two ablation variants (both causal) do not satisfy — they measure the same thing differently. Instruments must come from different epistemic traditions: e.g., causal + structural, or representational + behavioral.
  2. The instruments agree on component membership. Jaccard ≥ 0.5 is a reasonable pass threshold for circuits of ≤ 20 components.
  3. Disagreement is treated as a finding. Jaccard ≈ 0 must be reported as a primary result and investigated — not silently resolved by choosing the preferred instrument.

The weight-circuit (structural analysis) and EAP-circuit (attribution patching) have Jaccard ≈ 0. This is one of the project’s most informative findings. It means one of:

  • One instrument has a methodological flaw explaining its output;
  • The two instruments are sensitive to different real properties (structurally similar but causally inactive components vs. causally active but structurally atypical ones);
  • The construct “SVA circuit” is underspecified and instruments track different things under the same label.

All three interpretations are scientifically important. Convergent validity is failed; the disagreement is the finding.

Instrument coverage for convergent validity

Section titled “Instrument coverage for convergent validity”
InstrumentEvidence family
Weight classifier (B03)Structural
EAP attribution patching (A01)Causal
DAS-IIA (A02)Representational
Activation patching (A01)Causal
LLC per component (B04)Structural
Bootstrap stability (F01)Measurement

Any two instruments from different rows constitute a valid test.

  • List every instrument used and its evidence family.
  • For each pair from different families: Jaccard similarity.
  • If Jaccard ≥ 0.5: note which components appear in all instruments.
  • If Jaccard < 0.5: report disagreement explicitly, list components unique to each instrument, state which interpretation is most consistent with the data.