Model Governance

A deployable model should have a version, named owner, intended use, training or reference data lineage, feature definition, evaluation record and release state. Predictions are linked to the model version that produced them so later review does not depend on whichever model is current.

Candidate models are compared with transparent baselines and evaluated on held-out or time-appropriate data. Forecast error, coverage, stability and relevant subgroup performance are reviewed together; a single aggregate score is not treated as sufficient evidence.

Production models are monitored for data drift, prediction drift, failures and material degradation against their validation baseline. A new version is released separately rather than silently replacing its predecessor, and a problematic release can be disabled or rolled back while its outputs remain attributable.

Models support analysis; they do not carry institutional accountability. High-impact interpretations require human review before external use. AIC models must not be used as the sole basis for decisions that determine an individual's eligibility, employment, credit, healthcare, policing, migration status or access to essential services.

AIC has model-version, prediction, observation and drift-event records in its platform architecture. Formal independent model validation and a standing model-risk committee are not represented as complete. Institutional deployments should set their approval roles, thresholds and review cadence in the engagement plan.