Bay lights on, dies cold, and the inspection AI starts scrapping good parts. The model never got worse. The inputs did. Our Inspection Trust Gate demo stages exactly that morning. The marker reads "SHIFT CHANGE 06:00 - cold dies, bay lights on," and 12 good parts arrive with glare, defocus, and thermal cast. They are real MVTec AD metal_nut photographs; the drift is honest, labeled image corruption. The naive baseline (same defect model, no envelope check) false-rejects those good parts. The Trust Gate auto-scraps zero. Not by out-detecting anything: it first checks each image against the model's validated envelope, measured over eight physical capture signals. When the envelope monitor goes red, the gate stops trusting the model entirely and routes every drifted part to HOLD for human review instead of firing the reject actuator (a simulated Allen-Bradley adapter, labeled as such). Every decision writes a full audit lineage record: model version, dataset hash, OOD score, the rules that fired, latency against the 750 ms stroke window. 🏭 On that MVTec metal_nut held-out split, the no-envelope baseline false-rejected 95.5 to 100% of drifted good parts per drift family. The gate auto-scrapped 0%. Defective parts injected during the drift were still caught or escalated, 93 of 93. Demo measurements on a research benchmark, never open-world guarantees. Even a perfect defect model is only valid on inputs inside its validated envelope. The fix is not a better model. If your line runs edge inspection and false rejects spike on cold mornings or under new lighting, we would genuinely like to hear how you catch drift before the actuator fires. The problem is industry-wide and the answers will be too. #MachineVision #EdgeAI #QualityEngineering #SmartManufacturing #AIGovernance