Skip to content
Validity typeInterpretive
Pass conditionA specific description-mode tag is stated explicitly in the verdict
Evidence familyN/A (criterion is about claim structure)
Minimum reportingThe mode tag, stated verbatim, in the verdict section of any published claim
Common failure modePublishing a result without a declared mode tag; leaving the claim’s scope implicit

Level declaration is the simplest interpretive criterion: before any claim can be evaluated for interpretive validity, it must declare what level of description it is making. Without a declared level, there is no standard against which to measure evidence–claim fit.

The seven description modes:

ModeMeaningMinimum validity requirements
[computational]States what problem the system solves, without mechanismBehavioral evidence on a stated distribution
[algorithmic]Names the steps and their orderCausal evidence for each step and for the ordering
[representational]Claims a variable is encoded at a componentBaseline-separated IIA or equivalent
[implementational-topographic]Locates the computation in named componentsNecessity, with a matched-set control
[implementational-connectomic]Claims a specific edge structure between componentsEdge-level causal evidence, not node-level
[implementational-functional]Claims a component’s weights implement the computationStructural plausibility plus causal support
[implementational-activation]Claims a specific activation geometry carries the computationGeometry measured and separated from a baseline

A verdict without a declared mode is not a verdict — it is a measurement with a story attached.

The most common interpretive failure in MI is implicit level inflation: a paper establishes [representational] evidence (high IIA) and implicitly claims an [algorithmic] account without the additional evidence that requires. Level declaration forces the implicit claim to be explicit, where it can be evaluated.

Every verdict must name one of the seven modes, followed immediately by the scope restriction:

Verdict: [implementational-topographic] for L8.MLP as a primary SVA locus in GPT-2 Small on the Linzen et al. prompt distribution. No [algorithmic] claim is made — the operation L8.MLP performs is not established.”

Level declaration (V1) is the prerequisite for level–evidence match (V2). You cannot check whether the evidence licenses the claimed level until the claimed level is stated.