- A vision model that scores 97% in the lab can hit a 14% false-reject rate on the press. The model didn't degrade. The inputs left the distribution it was validated on. The fix is not a better model. It's knowing when not to trust the one you have 🧵
- The physics does the drifting: overhead bay-light glare shifting with stroke angle, lubricant pooling on warm vs cold dies, the first 50 parts of a shift before thermal equilibrium. That stamping example (200-ton press, 40 strokes/min) is from our research, 2026.
- Even a perfect defect model is only valid on inputs inside its validated envelope. Drift is an input failure, not a model failure. So the demo we built, the Inspection Trust Gate, is not a better detector. It sits between the vision model and the PLC reject actuator.
- The core is an EnvelopeDetector: Mahalanobis distance over eight physical capture signals (exposure, contrast, dynamic range, focus, high-frequency detail, thermal cast, saturation, glare fraction), fit on 220 known-good images. One question: is this image inside the envelope?
- Out of envelope means no verdict is trusted, however confident the model sounds. The part routes to HOLD for a human. What actuates is decided by a deterministic gate in plain code, on a hard 750 ms stroke-window budget. No LLM near the actuator. Agents advise, code decides.
- The number that carries the thesis: envelope AUROC 1.000 on a drift family (underexposure) the detector was never tuned against, on the held-out MVTec metal_nut test split. Without a held-out family, an envelope gate is circular: it only catches the drift you built it to see.
- Then the scripted shift hits "SHIFT CHANGE 06:00 - cold dies, bay lights on." 12 real good parts arrive corrupted (glare, defocus, thermal cast: synthetic drift on real MVTec photos, labeled in-app). The envelope monitor goes red on every one of them.
- Same images, both paths, on the MVTec metal_nut held-out split. The naive no-envelope baseline false-rejects 95.5 to 100% of drifted good parts per family (avg 98.9%). The gate auto-scraps 0, holds all. Full-strength corruptions: the claim is the direction, not the decimal.
- Defects don't hide under the drift: 93/93 defective parts in the held-out split were caught or escalated, never auto-passed. The detector itself is a deliberate stand-in (kNN over texture, PatchCore-lite). In production it's your NVIDIA Metropolis, Cognex, or custom model.
- Question for the OOD crowd: would you gate a reject actuator on Mahalanobis distance over physical capture signals, or insist on feature-space OOD from the model backbone itself? Where should the envelope live? #EdgeAI #OODDetection #MachineVision #SmartManufacturing
- A demo, not a deployment: the PLC and MES adapters are stubs that log what they would do; every decision writes an exportable audit lineage record. If you run inspection on a real line, we'd like to compare notes: https://veriprajna.com/demos/edge-ai-manufacturing-inspection
Published on X · July 20, 2026
On social media
See this post on its original platform
In our archive