Criterion I6 — Double Dissociation

Validity type	Internal
Pass condition	Two interventions cross: each breaks what the other spares
Evidence family	Causal
Minimum reporting	Both arms of the dissociation, both tasks, effect sizes for all four cells
Common failure mode	Showing necessity for one task without testing the converse

What this criterion requires

A double dissociation is a crossed design from neuropsychology (Shallice, 1988). It requires two interventions and two tasks: intervention A breaks task 1 but spares task 2; intervention B breaks task 2 but spares task 1. This rules out the possibility that either intervention is simply damaging the system generally.

Each half of the crossed design is already scored elsewhere: I1 covers the necessity arm and I4 the sparing arm. I6 scores only whether the two cross. The crossed design licenses a functional separation rather than a modular one (Plaut, 1995).

Satisfied when:

1. Two interventions are identified. Circuit A and circuit B, or component A and component B.
2. Two tasks are identified. Task 1 and task 2, matched for difficulty.
3. The crossed pattern holds. Ablating A breaks task 1 but not task 2; ablating B breaks task 2 but not task 1.

MI example

Across the sixteen audited claims, double dissociation caps every claim that reaches Mechanistically Supported. The field rarely attempts it. The IOI circuit has been shown necessary for IOI (ablation degrades IOI performance), but the converse — ablating a different circuit to show it breaks a different task while sparing IOI — is almost never tested. One exception: Feucht et al. (2025) achieved a double dissociation for induction heads, three years after the origin paper.

Connection to the chain

Required for Triangulated tier. Without a double dissociation, the mechanism could be a general-purpose component rather than a task-specific one. I6 is the single most common ceiling in the audited claims.