Skip to content

The five validity types are not a checklist to be completed in any order. They have a dependency structure: some types cannot be satisfied without prior work on others.

The arrows mean: the downstream type presupposes that the upstream type has been addressed. A gap upstream produces a hole in the downstream verdict even when the downstream evidence looks complete.

A circuit claim begins with a construct: the thing being claimed. If the construct is not defined independently of the components found by the discovery procedure, nothing downstream is interpretable.

The failure mode is construct conflation: the circuit is named for a behavior, and the name is then treated as if it described a coherent computational concept. Until the construct is defined clearly enough to be falsified — until there is a stated numerical threshold that would count as disconfirmation — the claim has no determinate content.

Practical consequence: The construct definition should be written before instruments are chosen.

Every number that internal validity rests on is produced by an instrument. If the instrument is unreliable, poorly calibrated, or lacking construct coverage, then the internal validity evidence is contaminated at the source.

The most common version of this failure: an IIA score of 0.48 is reported as evidence for a causal relationship, but the random-vector baseline is 0.44. The internal validity claim is built on a measurement that has not established baseline separation.

Practical consequence: Baseline separation, reliability, and sensitivity should be established before interpreting internal validity evidence.

External validity asks whether the claim generalizes. But “generalizes” is only meaningful relative to a well-established local result. If the internal validity case has not been made, a positive cross-architecture result is noise, not generalization.

Practical consequence: The external validity campaign (cross-task, cross-scale, cross-architecture) should begin after the internal validity case has reached at least Mechanistically supported tier.

Interpretive validity is applied last — to the verdict, not the evidence

Section titled “Interpretive validity is applied last — to the verdict, not the evidence”

Interpretive validity audits the verdict itself rather than the evidence beneath it. The question is whether the description-mode tag in the verdict matches what the evidence actually licenses. A finding can satisfy internal, external, and measurement validity and still fail interpretive validity if the verdict claims an algorithmic characterization that the evidence only supports at the implementational level.

Practical consequence: Write the verdict first at the lowest licensed mode tag. Upgrade only when the evidence explicitly licenses it.

TypePresupposesGated by
ConstructNothing
MeasurementA defined constructConstruct
InternalA calibrated instrumentMeasurement
ExternalAn established local resultInternal
InterpretiveA fully assembled verdictAll four above
PatternWhat it looks likeViolation
Circular constructCircuit named for what instruments foundConstruct: no independent definition
Uncalibrated IIAIIA 0.48 without random-vector baselineMeasurement gates internal
Single-seed generalizationReported as robust from one seedMeasurement (reliability) gates internal
Cross-arch before localCross-model result before within-model necessityInternal gates external
Algorithmic tag from ablation”implements SVA” from ablation aloneInterpretive: tag not licensed