Skip to content

The framework paper’s appendix scorecards call this criterion cross-model recurrence. It is the same criterion; this page keeps the name used in the criterion table, and the URL keeps the earlier slug so existing links resolve.

Criterion E4 — Cross-Model Generalization

Section titled “Criterion E4 — Cross-Model Generalization”
Validity typeExternal
Pass conditionThe corresponding causal organization recurs across independently trained models under a declared correspondence criterion
Evidence familyCausal, Structural
Minimum reportingModels tested, correspondence criterion, whether the causal organization (not just the behavior) recurs
Common failure modeObserving the same behavior in a second model and calling it the same mechanism

Cross-model generalization asks whether the mechanism appears in other models — not just the behavior, but the causal organization behind it. Different models can produce the same input-output relation through different internal mechanisms, so behavioral agreement alone does not establish generalization.

Satisfied when:

  1. A correspondence criterion is declared. What counts as “the same mechanism” across models — same heads, same subspace alignment, same causal graph structure.
  2. The causal organization is tested in a second model. Not just behavioral output, but the internal structure — which components, which connections, which information flow.
  3. The correspondence holds or fails transparently. If the mechanism recurs under one criterion but not another, both are reported.

Knowledge neurons claims “can be easily generalized” to other models, but the study examines only one model. Cross-model generalization is untested. This is a scope-creep failure (V5) compounded by the absence of E4 evidence.

Required for Triangulated where the claim asserts reach beyond the systems tested; where it does not, prompt generalization (E2) carries the gate instead. Without cross-model generalization, the mechanism may be an idiosyncrasy of one training run rather than a general computational strategy. Steel (2008): recurrence alone supports limited induction; stronger extrapolation requires evidence that the causally relevant process is preserved.