Distribution Drift Bound

Foundation models are validated once and deployed indefinitely. Between validation and failure, their output distributions drift — silently, continuously, and without governance evidence. Static benchmarks provide no operational assurance. Periodic evaluations sample behavior at checkpoints but miss the drift between them. Observability dashboards display metrics but offer no cryptographic integrity, no formal calibration, and no bridge to regulatory compliance or commercial insurance.Clause AI-6 specifies the Distribution Drift Bound, the fourth mandatory invariant in the Model Attestation Interface (MAI-1) within the Auburn Governance Stack. The specification requires that the KL divergence between a model's production output distribution and its cryptographically sealed baseline distribution remain below a certified threshold at every attested measurement point. This document provides the complete technical infrastructure to operationalize that requirement.