Criterion C4 — Discriminant Validity

Validity type	Construct
Pass condition	The measure distinguishes this construct from neighboring constructs that share surface features
Evidence family	Measurement-theoretic
Minimum reporting	Named neighboring construct, method used to separate them, separation statistic
Common failure mode	Presupposing separation without testing it

What this criterion requires

Discriminant validity establishes that the construct under study is distinct from related constructs that share overlapping surface features. Two constructs that cannot be empirically separated are one construct with two names.

Satisfied when:

1. A neighboring construct is named. The claim identifies what the mechanism is not, not only what it is.
2. The measure separates the two. The metric that scores the target construct gives a different score for the neighbor, and the difference exceeds baseline variance.
3. The separation is not an artifact of the metric. A metric designed to detect X will score higher on X-labeled data by construction. The test must be capable of failing.

MI example

Gender bias circuits claim to localize a “bias” mechanism. But gender bias and gender competence (knowing which pronoun applies) share the same representations. Without a test showing the circuit carries bias but not competence — or vice versa — the two constructs are not separated, and the claim lacks discriminant validity. The gender bias case study is scored C4 Untested.

Connection to the chain

Discriminant validity is required for Triangulated tier. Without it, convergent validity (C3) is uninterpretable: multiple methods may converge on a construct that is not the one the claim names.