System architecture

flowchart LR
    A["Canonical or legacy CSV"] --> B["Python validation and conversion"]
    B --> C["2016-only imputation and scaling"]
    C --> D["Shared 2017 evaluation dates"]
    D --> E["Seasonal naive"]
    D --> F["Rolling ARIMA"]
    D --> G["32-unit LSTM"]
    D --> H["Versioned JSON/CSV job"]
    H --> I["MATLAB R2024b reference EFS"]
    H --> N["Independent Python IT2 EFS"]
    I --> J["Backend-identified predictions, model, rules, activations"]
    N --> J
    E --> K["Audited run artifact"]
    F --> K
    G --> K
    J --> K
    K --> L["CLI and Streamlit"]
    K --> M["Static GitHub Pages evidence"]

Backend boundary

Python owns the data contract, leakage controls, evaluation dates, baseline models, metric aggregation, and artifact manifest. MATLAB owns interval type-2 FIS initialization, three-stage optimization, inference, and rule explanation. The boundary is a versioned JSON job plus CSV inputs and outputs; no MATLAB Engine binding is required.

The optional Python IT2 backend consumes the same job and feature contract and emits the same evidence classes, using model.npz plus hash-bearing model.json instead of model.mat. Artifacts explicitly identify matlab or python-it2; cross-backend metrics are never treated as repeat seeds of one implementation.

For a normal workstation, Python launches matlab -batch. Public GitHub runners use the same job files but execute them through the official MathWorks Action because its temporary public project license is scoped to that action. Cached job output is accepted only after input hashes, features, seed, and EFS configuration match.

GitHub Pages is deliberately static. It presents checked-in evidence and never executes MATLAB or accepts private datasets.