Submission and source-alignment checklist

Before an empirical application, retain:

1. an immutable source deposit, version, configuration, and checksums computed from actual deposited bytes;

2. the complete rule functions, boundary conventions, rule priors, switch law, and optional label modifications;

3. the precise information supplied before and after prediction;

4. fixed versus online-updated checkpoint status;

5. sequence, trial, segment, and window definitions;

6. all metric numerators, denominators, exposures, and censoring;

7. current weights, masks, references, scoring interfaces, optimizer policy, and named streams;

8. executed candidate identities and complete intervention budgets;

9. explicit matching statuses and surrogate assumptions;

10. calibration and fresh assessment separation;

11. independent training replicates when procedure-level claims are made;

12. a separate construct-validation dossier before cognitive or clinical interpretation.

A source title, a notebook heading, a mutable repository root, or a numerical summary is not a complete source-alignment certificate. A mathematical proof, a software execution, and a biological correspondence are different forms of evidence.