The Wrong Scale of Explanation: Effective Structure, Distributed Intelligence, and the Category Error of Microscopic Localization

Abstract

The Wrong Scale of Explanation: Effective Structure, Distributed Intelligence, and the Category Error of Microscopic Localization Civilization Physics - Series: Effective Structure / AI Interpretability This article argues that neural-network explanation is often distorted by a hidden demand for microscopic localization. Interpretability research productively asks where concepts, features, circuits, and behaviors appear inside a model, but that question can become scale-mismatched when the phenomenon is distributed, relational, redundant, contextual, or multiply realizable. A precise local answer may still miss the explanatory object. The central claim is that explanation should match the scale at which the phenomenon remains stable. Effective structure names a relational organization that compresses microscopic variation while preserving prediction, intervention, counterfactual stability, recurrence, and explanatory economy. Physical realization is required, but one-to-one microscopic localization must be demonstrated rather than assumed. The article develops this argument through several linked mechanisms: The Microscopic Localization Requirement turns a useful research strategy into a universal but often false standard of explanation. The scale-matching principle asks researchers to choose the coarsest effective description that preserves the causal and predictive distinctions required by the question. Effective structure is defined through context-conditioned relations, topologies, state partitions, algorithms, constraints, representational geometries, or dynamical regimes. Load-bearingness is tested counterfactually: a distinction matters when changing it changes the task-relevant future. Mechanistic interpretability remains essential for implementation, causal control, failure discovery, and cross-scale linkage, but it should not monopolize explanatory legitimacy. The Question-Scale Audit requires researchers to name the phenomenon, identify its scale of recurrence, define equivalence, declare interventions, constrain abstraction, and state falsifiers before choosing tools. This article reframes interpretability as a multi-scale science of distributed intelligence. Behavioral evidence, effective structural analysis, mechanistic tracing, and training-level explanation should constrain one another rather than compete for exclusive authority. The final principle is methodological: requiring every effective structure to appear as a localized internal object confuses implementation with explanation, while unconstrained high-level mapping turns explanation into storytelling. Keywords: effective structure, scale of explanation, distributed intelligence, mechanistic interpretability, microscopic localization, causal abstraction, neural representation, emergence, sparse autoencoders, load-bearing structure, scale-matching principle, Question-Scale Audit, AI interpretability, multi-scale explanation