BLUM: Brain-LLM Unified Model — Source Code and LLM-side Derived Data

Source code and LLM-side derived data accompanying Fridriksson et al., "Stroke Lesions as a Rosetta Stone for Language Model Interpretability" (submitted to Patterns, 2026; arXiv:2602.04074). The Brain-LLM Unified Model (BLUM) framework uses human lesion-symptom mapping in chronic post-stroke aphasia as an external reference structure for evaluating large-language-model perturbation effects. This deposit contains the full Phase 1–5 analysis pipeline (data preparation, ridge-regression training with leave-one-out cross-validation, multiplicative-Gaussian perturbation of transformer layers, six-category error classification, projection of LLM error profiles into a 24-ROI human lesion space, top-k matching against a 2,000-sample random baseline, and dual-stream permutation testing), the Streamlit step-through GUI, the JHU-atlas ROI definitions used as the analytic target space, and the LLM-side derived data: full LLM error-profile matrices for all retained perturbation conditions (1,986 PNT and 2,227 WAB-R). The human-derived data underlying the symptom-to-lesion model training and the LLM-vs-human matching analyses (per-participant behavioural error proportions, ROI-level lesion loads, item-level behavioural transcripts, and pre-trained symptom-to-lesion model objects) remain governed by the lab's Data Use Agreement and are available from the corresponding author on reasonable academic request. Raw T1/T2 MRI and lesion masks for the same cohort are publicly available via OpenNeuro under that platform's standard sharing terms. The BLUM methodology is the subject of U.S. Patent US-10916348-B2 and U.S. Provisional Application No. 63/974,622. The source code is released under the Apache License 2.0, which includes an explicit patent-license grant: users of the released code may use it for academic and commercial purposes covered by the released implementation. Researchers wishing to apply the broader patented methodology beyond the released code should contact the corresponding author.