Methodological Framework and Data Management Plan for Mechanistic Interpretability Audits – Slovak Open Science Forum 2026

This deposit contains the methodological framework, Data Management Plan (DMP), and public engagement strategy prepared for the Slovak Open Science Forum 2026. It outlines a reproducible and secure architecture for auditing open-weight neural networks using mechanistic interpretability techniques. The framework establishes protocols for deploying transformer models within local, air-gapped sandboxes to conduct feature probing and sparse autoencoder (SAE) analysis. In strict adherence to FAIR data principles and European Open Science Cloud (EOSC) guidelines, the DMP ensures that raw activation tensors and vulnerability exploits remain securely restricted. Conversely, anonymized activation maps, diagnostic codebases, and synthetic prompt suites are structured for open-access deposition. Additionally, the included public engagement strategy bridges these technical workflows with national open science infrastructure goals, detailing integration pathways for preprint platforms, the OpenAIRE Graph, and digital humanities repositories. Keywords: Mechanistic Interpretability, AI Safety, Open Science, FAIR Principles, Data Management Plan, Neural Networks, Sparse Autoencoders