Model-Drift-Triggered Safe-State
Objective
Ensure safe-state logic accounts for gradual AI model drift, not only discrete hardware/software faults.
Control / requirement
Monitoring for gradual divergence between expected and observed model behavior; defined safe-state or degraded-mode response to sustained drift, distinct from discrete-fault triggers.
Applicability
Systems using learned models whose behavior can drift from validated performance over time or operating conditions.
Expected evidence
Drift-detection methodology; drift-triggered transition test records. [T3]
Assurance expectation
Test evidence demonstrating drift detection at a meaningful threshold before drift becomes unsafe, not only after failure.
Dependencies
PAI-SF-SEN-002 (Domain 1, sensor calibration drift) for sensor-side drift; distinct from model-side drift addressed here.
Exclusions
Not applicable to systems using only deterministic, non- learned control logic with no drift-capable component.
Maturity / conformance relevance
Expected at Operational and High-Assurance levels for learned-model systems.
Ecosystem relationship
This is the clearest AI-specific gap PAI-SF™ fills relative to IEC 61508/ISO 26262, which model discrete faults, not statistical drift — the framework's core justification made concrete at control level.