Scientific Theory of a Black-Box: A Life Cycle-Scale XAI Framework Based on Constructive Empiricism

XAI needs a persistent record, not just isolated explanations: The paper identifies a critical gap in current explainable AI practice—while many XAI methods exist to answer specific questions about AI models, there's no standard way to maintain a comprehensive, auditable record of explanations throughout a model's entire operational life cycle.

Three core principles should guide AI documentation: The framework proposes that any explanation system should meet three obligations: (1) accurately reflect all observed model behavior, (2) be updatable when new observations reveal gaps in understanding, and (3) maintain transparent documentation of all assumptions and decisions made during the explanation process.

Explanations should be queryable records, not one-off outputs: Rather than generating explanations on-demand each time someone asks a question, the paper advocates for maintaining a living document of model behavior that stakeholders can query for different purposes—similar to how scientific theories are consulted to answer different questions rather than being recreated each time.

Rule-based surrogates can track model behavior over time: The proof-of-concept demonstrates that simpler, interpretable rule-based models can be continuously updated to match a complex AI system's behavior as it encounters new data, providing an auditable trail of how the model's decision-making has evolved.

Auditability requires explicit documentation of construction choices: The framework emphasizes that responsible AI requires transparent records not just of what the model does, but how explanations were built, what assumptions were made, and how the explanation system will adapt—enabling third-party assessment and accountability.