This preprint proposes a governance-oriented framework for the retrain-or-recalibrate decision in high-stakes machine learning systems. Building on prior work on deployment reliability under temporal distribution shift, the paper reframes model maintenance as an auditable organisational decision involving reliability volatility, operational cost, carbon cost, and lifecycle accountability. The paper introduces design requirements, a maturity model, and an audit-oriented decision structure for monitoring discrimination, calibration, distributional drift, and explanation stability over time. The contribution is positioned as a design science artefact for responsible and sustainable AI lifecycle governance.
Rahman et al. (Mon,) studied this question.