New Protocol Improves AI Interpretability and Reusability

Mechanistic interpretability aims to understand how neural networks function at a component level, generating detailed analyses of what individual parts encode and how they interact. However, the outputs of these analyses, such as selectivity tables or circuit diagrams, are often confined to individual research notebooks, making them difficult to reuse, query programmatically, or apply directly for auditing or intervention in downstream tasks. This research identifies the representation layer, which bridges these analyses and their practical application, as a critical bottleneck. To address this, the authors propose Manifestation Units, a structured, typed tuple protocol (E, S, R, D, G) designed to organize per-component statistics. For transformer architectures, this protocol is extended with attention-head primitives (T). This framework allows for automatic population of structured fields and querying through hybrid retrieval methods. The protocol was instantiated across various models, including generative vision (beta-VAE), discriminative vision (CNN), and language (GPT-2). Findings indicate that this typed structure substantially outperforms unstructured baselines in retrieval tasks. Furthermore, CNN filters retrieved using the schema satisfied causal sufficiency and necessity criteria under controlled conditions. The schema also effectively absorbed attention-head primitives without modification and revealed a core two-field structure (S+R) as irreducible, with other fields being either redundant or interfering. This work presents a foundational schema infrastructure for advancing mechanistic interpretability.