Criterion E6 — Cross-Architecture Generalization

Validity type	External
Pass condition	The mechanism appears in at least one other model family (not just a different checkpoint or size of the same model)
Evidence family	Structural, Behavioral
Minimum reporting	Model family tested; matching criterion used; structural alignment score or behavioral replication result; null expectation
Common failure mode	Claiming generalization after testing only different sizes of the same model family

What this criterion requires

Cross-architecture generalization is the strongest form of external validity. The mechanism must be found in a model family with different architecture, training data, or training procedure.

Satisfied when:

1. A second model family is tested. GPT-2 variants (Small, Medium, Large, XL) are the same family. GPT-2 Small + Pythia-160M is a cross-family test. GPT-2 Small + Gemma-2B is a stronger cross-family test (architectural differences: grouped-query attention, RMSNorm, SwiGLU).

2. An explicit matching criterion is stated. How are circuit components identified as “the same” across models? Acceptable criteria: cosine similarity of W_OV decoder directions ≥ 0.7, weight classifier F1 on cross-model component identification, IIA transfer on the same task.

3. The result meets the matching criterion with a stated null expectation (e.g., expected cosine between random heads from different models).

Current project status

Gemma-2B and Qwen-0.5B initial investigations are complete. This constitutes the beginning of the cross-architecture generalization campaign. Current status: Proposed [transportable] — structural alignment checks initiated, causal transport (cross-model IIA or ablation) not yet completed.

Upgrade path to Causally suggestive: Run ablation or DAS-IIA on the same task in at least one of these models and show above-baseline performance.

Why this matters for publishability

A circuit finding that generalizes across model families is substantially more significant than one found in a single model. The IOI circuit has been partially validated across models; the SVA circuit has not. Demonstrating cross-architecture generalization for any circuit in this project would upgrade from Mechanistically supported to Triangulated (given other criteria met) — a publishable contribution to the cross-model circuit literature.

Minimum reporting rule

• Second model family tested and its architectural differences from the discovery model.
• Matching criterion stated explicitly.
• Structural alignment score or behavioral replication result with null comparison.
• Distinguish “same family, different size” (robustness, E5) from “different family” (cross-architecture, E6).