Eidon – Neural State Manifold and Modelling System: A Unified Architecture for Harmonised EEG/MEG Feature Extraction and Multi-Framework Modelling

Open EEG and MEG repositories now contain thousands of hours of neural recordings across diverse tasks, subjects, and acquisition systems. However, large-scale comparative analysis and model-based inference remain difficult due to heterogeneous data formats, bespoke preprocessing pipelines, inconsistent metadata, and non-standardised feature definitions. These limitations impede reproducibility and restrict cumulative evaluation of theoretical and applied claims about neural dynamics, arousal, cognition, and behaviour. This paper specifies the architecture of Eidon: Neural State Manifold and Modelling System, a unified, declaratively configurable EEG/MEG analysis and modelling framework designed to produce comparable feature assets and modelling artefacts across heterogeneous sources. The framework is designed to: (i) harmonise heterogeneous EEG/MEG inputs via a shared internal schema; (ii) apply fully parameterised preprocessing; (iii) compute a spatially referenced, cross-domain electrophysiological feature stack spanning spectral and cross-frequency measures, aperiodic (1/f) components, time-domain and ERP features, state-dynamics and complexity measures, ICA-derived features, connectivity matrices, and graph/network metrics; (iv) align neural features with behavioural and subject-level variables; and (v) support model-agnostic comparative modelling by enabling multiple model families to be fit and evaluated on shared, standardised neural and behavioural feature representations. Supported model families include classical statistical models, machine-learning approaches, probabilistic generative models, dynamical systems models, network and graph-based models, and quantum-inspired cognitive models, with optional extension modules available for additional specialised model classes where justified by a given use case. All stages are governed by a single configuration file specifying data sources, parameters, feature sets, model families, evaluation schemes, export targets, and optional run-level execution controls, enabling reproducibility, auditability, and transparent provenance. The paper has two aims. First, it specifies the system design, rationale, and module interfaces for an end-to-end automated EEG/MEG pipeline whose specification is defined independently of any particular implementation. Second, it defines the configuration semantics, output schemas, and provenance guarantees required to support reproducible cross-dataset feature extraction and comparative modelling across heterogeneous EEG/MEG sources.