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

Running one metric and claiming a circuit is not enough. Convergent validity requires triangulation across independent measurement traditions.

Satisfied when:

  1. ≥3 metrics from different evidence families have been run. Two ablation variants (both causal) do not satisfy — they measure the same thing differently. Families count as independent when the failure of one’s core assumption would not automatically invalidate the other: e.g., causal + structural + representational.
  2. The metrics agree on component membership. Jaccard ≥ 0.5 is a reasonable pass threshold for circuits of ≤ 20 components.
  3. Disagreement is treated as a finding. Low Jaccard overlap must be reported as a primary result and investigated — not silently resolved by choosing the preferred metric.

When two discovery methods from different evidence families nominate substantially different component sets (Jaccard < 0.5), the disagreement typically means one of:

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

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

Metric typeEvidence family
Attribution patchingCausal
DAS-IIARepresentational
Weight-space analysis (SVD, composition scores)Structural
Activation statistics (LLC, probe accuracy)Structural / Representational
Bootstrap stabilityMeasurement

Three metrics from different evidence families constitute the test at Triangulated; two establish a partial result.

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