The Drift You Cannot See Until It Costs You

Production teams validate a model against a frozen alias and assume the behavior holds. It does not. Between named versions, hosted models drift silently through safety fine-tuning, inference-engine changes, gateway routing, and system-prompt edits — documented in Anthropic's March–April 2026 postmortem and the GateScope gateway study — while cost, latency, and version dashboards stay green. This brief argues for behavioral fingerprinting: a fixed probe set, run on a schedule, compared to a baseline, with a layered alert and a dated trail.