The parts were fine. At the 06:00 shift change, cold dies, bay lights on, the inspection AI starts scrapping them anyway. A vision model that scores 97% in the lab can hit a 14% false-reject rate on a stamping press. Not because it got worse. Glare and thermal drift walked the images outside the conditions it was validated on. Drift is an input failure, not a model failure; a perfect model fails the same way. So we built the Inspection Trust Gate, a runnable demo of the missing layer. Before any verdict reaches the reject actuator, plain deterministic code asks the one question the model cannot answer about itself: is this image inside the validated envelope? Outside it, the part goes to HOLD for a human, stamped with an EU-AI-Act-ready audit lineage record. That opening scene is the demo's scripted shift (real MVTec AD metal_nut photographs, labeled synthetic glare, defocus, and thermal-cast drift, a simulated PLC actuator). Twelve good parts arrive after the shift change. The naive no-envelope baseline scraps all 12. The gate scraps 0 and holds every one for a person; the model's verdict never reaches the actuator. When your line drifts at shift change, does anything check the image before the verdict fires the actuator, or does the model's word go straight to the PLC? #MachineVision #IndustrialAutomation #EdgeAI #EUAIAct