The runtime trust layer between the vision model and the actuator

Your inspection AI does not have an accuracy problem. It has an envelope problem.

A 97% lab model can hit a 14% false-reject rate in production, not because the model got worse but because glare, cold dies, and shift changes walked the inputs out of the distribution it was validated on. The Inspection Trust Gate checks every image against the model's validated envelope, applies deterministic gates the model cannot override, holds anything out-of-envelope for human review, and writes a per-part audit lineage record. Agents advise, code decides.

Drifted good parts auto-scrapped by the gate

Naive baseline false-rejects 95.5 to 100% (MVTec metal_nut held-out split)

93/93

Defective parts under drift still caught or escalated

Never auto-passed (same held-out split)

1.00

Envelope AUROC on a held-out drift family

Never tuned against it, so the check is not circular (MVTec metal_nut held-out split)

This is a runnable demo. The PLC actuator and MES sink are simulated adapters that log what they would do, the part images are real MVTec AD metal_nut photographs, and the drift is honest image corruption applied to real good photos.

A 97% model in the lab. A 14% false-reject rate on the press.

The failure mode that stalls edge inspection deployments is an input failure, not a model failure.

The canonical case from our research is a 200-ton progressive die press running 40 strokes per minute: a vision model that hit 97% accuracy in validation collapses to a 14% false-reject rate in production. The model did not get worse. Overhead bay-light glare varies per stroke angle, lubricant pools differently on warm and cold dies, and the first 50 parts of a shift arrive before thermal equilibrium. The inputs left the distribution the model was validated on, and no model, at any accuracy, is trustworthy there.

The industry pattern says the model was never the project. Out-of-box AOI systems false-reject 5 to 15% of good parts, while well-tuned systems get under 2%: a calibration and data problem, not a model architecture problem. Integration is 60% of a deployment's timeline and model training 15%, and 84% of system integration projects fail or partially fail.

Regulation raises the stakes. The EU AI Act's high-risk obligations, which cover safety-critical quality decisions under Annex III, are fully applicable from August 2, 2026, with maximum fines of 35 million euros or 7% of global turnover for the most serious, prohibited-practice violations. A reject actuator that fires on an unvalidated model verdict, with no record of why, is exactly what that regime is designed to expose.

The fix is not a better model. Even a perfect model is only valid on inputs inside its validated envelope. What the line needs is a runtime layer that knows that envelope, refuses to actuate outside it, and can prove, part by part, what it decided and why.

How the Inspection Trust Gate works

A deterministic gate decides what actuates, not a model and not an LLM.

Every part runs the same pipeline:

image → physical + texture features → envelope check (OOD) → defect verdict → deterministic gate → actuate / hold → audit record

1. The validated envelope

An EnvelopeDetector computes Mahalanobis distance in physical-signal space (exposure, contrast, dynamic range, focus, high-frequency detail, thermal colour cast, saturation, glare fraction), fit on 220 train-good images. It answers one question: is this image inside the capture conditions the model was validated on? Out-of-envelope means no model output is trusted.

2. The deterministic gate

Plain code, outside any LLM, that the model cannot override: calibrated confidence thresholds fitted from the demo's own data (auto-pass below 0.948, auto-reject above 1.30) and a hard 750 ms stroke-window latency budget. Each part routes to AUTO_PASS, AUTO_REJECT, or HOLD, which sends it to a human escalation queue.

3. The audit lineage

Every decision writes a JSONL record: part id, station, model id and version, dataset hash, defect confidence, OOD score, physical signals, the gate rules that fired, latency against budget, actuation and MES logs, and the risk tag high-risk:quality-gate (EU AI Act Annex III, eff. 2026-08-02). One click exports the shift.

The defect model is deliberately not the product

The demo's DefectDetector is a kNN distance to train-good parts in texture space (PatchCore-lite), standing in for your NVIDIA Metropolis, Cognex, or custom model behind a fixed interface. It is a real working stand-in (AUROC 0.845 on clean good versus defect), reported honestly, and the product claim never rests on it. The durable value is the layer around it: drift-gating, provenance, and governed actuation hold at any model accuracy, so the layer survives every model upgrade.

Agents advise, code decides

When parts pile up in the HOLD queue, an agentic Drift Triage pair explains why: a diagnosis agent reads the ranked physical-signal deviations of the held parts and proposes a root-cause hypothesis with a recommended action, and a critic agent verifies that hypothesis against the numeric evidence, downgrading it to manual investigation if the cited signal is not actually the dominant deviation. The agents are built on Pydantic AI and are provider-swappable; with no API key configured, triage falls back to a deterministic template so the whole demo runs offline. Either way, the agents only advise. The gate has already decided.

The shift change, worked end to end

A scripted shift on station Line 3, replayed over real MVTec AD metal_nut held-out test parts. Every image below is a screenshot of the running app.

Normal operation: clean parts pass in tens of milliseconds

Before the drift, clean good parts auto-pass with time to spare: the shipped audit log shows part test-good-288 at 37.7 ms and test-good-289 at 24.4 ms against the 750 ms stroke-window budget, each with a full lineage record. The dashboard shows the live decision pipeline per part, with each stage's measured wall-clock and the evidence behind its output.

The Inspection Trust Gate dashboard during normal operation: a clean part with a PASS verdict, the live decision pipeline showing per-stage milliseconds, and the envelope monitor nominal. — A clean part auto-passing: PASS verdict, per-stage timings, envelope monitor nominal.

06:00, shift change: the inputs leave the envelope

Then the scripted marker fires: SHIFT CHANGE 06:00, cold dies, bay lights on. Twelve held-out good parts arrive corrupted with glare, defocus, and thermal cast (honest image corruption applied to real good photos, labeled as such in the app). The envelope monitor goes red. The gate holds every one of them for human review instead of letting the actuator fire on a verdict the model was never validated to give.

The drift moment: SHIFT CHANGE 06:00 banner, a HOLD verdict reading needs proof, an out-of-distribution breach ring around the part, and a note that a naive AOI would have rejected it. — The drift moment: HOLD instead of a false reject. The naive AOI path would have scrapped this good part.

The envelope breach explained signal by signal: brightness at plus 7.6 sigma contributing 87.1% of the Mahalanobis distance. — Why the envelope tripped: brightness at +7.6 sigma, 87.1% of the distance. Evidence, not a black-box flag.

The same shift, measured on both paths

The comparison panel runs the identical parts through the naive no-envelope baseline and the Trust Gate. The baseline false-rejects the drifted good parts; the gate auto-scraps 0 and holds them, driving the false-reject KPI from 57.1% to 0% in this scripted shift. A projection panel extrapolates the measured naive false-reject rate to a full shift (40 strokes per minute over 8 hours is 19,200 parts) at $2.42 per part, a sourced order-of-magnitude scrap cost (cookie-manufacturer case: $94K annual savings from an 8.7% scrap-waste reduction). The dollar figure is always a labeled projection, never a measured saving.

The payoff panel: naive baseline 12 good parts scrapped versus Trust Gate 0, the escalation queue of held parts, the false-reject KPI falling from 57.1% to 0%, and a labeled scrap-cost projection of about $26.5k. — Naive 12 good parts scrapped, gate 0. The held parts wait in the escalation queue; the dollar counter is a labeled projection.

A real defect auto-rejects, honestly annotated

Drift-gating would be worthless if defects hid under it. A real gross structural defect (MVTec class flip, part test-flip-264) auto-rejects, and the simulated actuator logs REJECT actuated in 25 ms against the 750 ms budget. The geometry drill-in states that nothing was localised on this part: the reject rests on texture confidence alone, and the zone rule fell back to its centre default. Showing that text is deliberate. The demo displays the rule abstaining rather than pretending. Across the whole held-out split, defective parts injected during drift are still caught or escalated, 93 of 93, never auto-passed.

The full decision trace of the auto-rejected real defect test-flip-264, with the simulated actuator log showing REJECT actuated in 25 ms against the 750 ms budget. — The auto-reject trace: a real defect, actuation logged by the simulated adapter.

The honesty drill-in on the auto-rejected part: no anomaly localised, the reject rests on texture confidence alone. — The disclosure, in the product: no anomaly localised, texture confidence alone.

The receipt, and the explanation

Click any part and the audit drawer shows its full lineage record: model id metalnut-defect-knn version v7, dataset hash ae95b5b533c8, OOD score, physical signals, the rules that fired, latency against budget, the actuation and MES logs, and the risk tag high-risk:quality-gate (EU AI Act Annex III, eff. 2026-08-02). Export Audit downloads the shift as inspection_audit.jsonl; audit completeness is 100% of decided parts on the held-out split. Run Drift Triage and the diagnosis agent proposes a root cause for the held parts while the critic agent verifies it against the numeric evidence before it earns a verified badge.

The per-part audit lineage JSON with model id and version, dataset hash, OOD score, gate rules fired, latency, and the EU AI Act Annex III high-risk tag. — The per-part lineage record, designed as filable evidence for a high-risk conformity file.

Drift Triage output: the diagnosis agent's root-cause hypothesis and recommended action, with the critic agent's verified badge. — Drift Triage: a diagnosis agent proposes, a critic agent verifies. Advisory only; the gate has already decided.

The numbers, with their exact scope

All measurements come from the demo's benchmark script on the MVTec AD metal_nut official held-out test split (220 train-good fit images; 22 good and 93 defective test parts), comparing the naive no-envelope baseline against the Trust Gate on the same images. The baseline false-rejects 95.5 to 100% of drifted good parts per drift family (average 98.9%); the gate auto-scraps 0.0% and holds 100% of them. The envelope detector scores AUROC 1.00 on a held-out drift family (underexposure) it was never tuned against. The corruptions are full-strength, so the baseline's collapse is near-total by construction: the honest claim is the direction, a clean-validated model collapses once inputs leave the envelope, not the exact percentage. These are demo measurements on a research benchmark, never open-world guarantees.

Naive AOI pipeline versus the Inspection Trust Gate

The same comparison the demo runs live, on the same parts.

Dimension	Naive AOI (no envelope check)	Inspection Trust Gate
Out-of-envelope inputs	Model verdict trusted anyway	No model output trusted; part HELD for review
Drifted good parts (glare, defocus, thermal cast)	95.5 to 100% false-rejected per drift family	0 auto-scrapped; 100% held (MVTec metal_nut held-out split)
Who fires the actuator	The model's verdict, directly	A deterministic gate the model cannot override
Latency accountability	None	Hard 750 ms stroke-window budget, measured per part
Audit trail	None	Per-part JSONL lineage record, exportable per shift
When the line drifts	Scrap piles up until someone notices	Escalation queue plus advisory Drift Triage (diagnosis agent, critic agent)

What this demo does not do

✓ It does not run a live line. The EtherNet/IP to Allen-Bradley ControlLogix reject actuator and the MES sink are simulated adapters that log what they would do, and the camera is a recorded image stream. Station Line 3 and the shift are scripted.
✓ It does not claim open-world guarantees. The 0.0% gate false-reject, the 1.00 envelope AUROC, and the 93/93 catch rate are measurements on the MVTec metal_nut held-out split under full-strength synthetic corruptions.
✓ It does not show real plant drift. The drift is honest image corruption (glare, defocus, thermal cast) applied to real MVTec photographs: real photos, synthetic drift, labeled as such in the app.
✓ It does not sell the defect model. The kNN stand-in exists so the layer around it can be shown; in production the model is yours, and we make no claim to out-detect AOI vendors.
✓ It does not ship production metrology. The geometric zone rule is a coarse stand-in that localises an anomaly on 33 of 93 held-out defects and changes the gate's outcome on 1 of 93; the UI discloses this per part.
✓ It does not report measured savings. The scrap-cost counter is a labeled projection at $2.42 per part, a sourced order-of-magnitude, and no customers, deployments, or case studies exist for this demo.

Questions buyers actually ask

Why is our vision system rejecting good parts after a shift change?

Usually because the inputs drifted, not because the model got worse. Bay-light glare, cold dies, and the first parts of a shift before thermal equilibrium move images outside the distribution the model was validated on, and no model is trustworthy on inputs outside its validated envelope. A 97% lab model hitting a 14% production false-reject rate on a stamping press is the canonical example. The durable fix is a gate that detects out-of-envelope inputs and holds those parts for review instead of actuating.

Do we have to replace our existing inspection model or AOI system?

No. The Inspection Trust Gate wraps your model behind a fixed interface; in the demo the defect detector is a deliberate stand-in for an NVIDIA Metropolis, Cognex, or custom engine. The product is the layer around the model: the envelope check, the deterministic gate, the escalation queue, and the audit lineage. That layer keeps earning its keep through every model upgrade, because the failure it prevents is an input failure, not a model failure.

Is it fast enough for a real stamping line?

The gate is plain code with a hard latency budget: the reject decision must land inside the 750 ms stroke window of a press running 40 strokes per minute. In the demo's shipped audit log, decisions land in tens of milliseconds, for example 24.4 ms and 37.7 ms against the 750 ms budget, and the dashboard reports the live p99. These are demo measurements, not a production guarantee, but the budget is enforced per part.

Does the EU AI Act apply to AI quality inspection, and what do we have to log?

Safety-critical quality decisions fall under the Act's high-risk obligations (Annex III), fully applicable from August 2, 2026, with maximum fines of 35 million euros or 7% of global turnover for the most serious, prohibited-practice violations. The demo writes a per-part lineage record: model id and version, dataset hash, OOD score, the rules that fired, latency against budget, actuation log, and the risk tag, exportable as JSONL. That record is designed to be filable evidence for a high-risk conformity file; it is not a certification.

Can the AI, or an LLM, fire the reject actuator?

No. Actuation is decided by a deterministic gate: fitted confidence thresholds and a latency budget in plain code that no model output and no LLM can override. The agentic part, Drift Triage, is a diagnosis agent plus a critic agent that explain why the line drifted; they advise and never actuate. With no API key configured, triage degrades to a deterministic fallback and the whole demo runs offline.

How do we know the drift detector is not just tuned to the demo?

That is the independence check we ran: the envelope detector scored AUROC 1.00 on a held-out drift family (underexposure) it was never tuned against, measured on the MVTec metal_nut held-out split. It is fit on 220 known-good images in physical-signal space, so it flags capture drift that moves those signals, not just the drift families we built. The corruptions are full-strength, so the separation is stark by construction; the honest claim is the direction, not the exact percentage.

Is this a live line or a demo?

A runnable demo that proves the mechanism. The EtherNet/IP to Allen-Bradley ControlLogix reject actuator and the MES sink are simulated adapters that log what they would do, and the camera is a recorded stream of real MVTec AD metal_nut photographs; the drift is honest image corruption applied to real good photos. The envelope check, the deterministic gate, the escalation queue, and the audit export are real and run exactly as shown, and every headline number is measured on the MVTec metal_nut held-out test split.

Technical Research

The research behind this demo — the architecture, the verification design, and the enterprise blueprint.

Full solution

Explore the Edge AI for Manufacturing Quality Inspection solution →

Recovering or hardening an edge inspection deployment?

The trust layer between the model and the actuator is the hard part. We build it.

If your team is wrestling with false rejects after every shift change, or with what the EU AI Act's high-risk obligations mean for a line that decides at 40 strokes per minute, we would genuinely like to hear how you are thinking about it. The problem is industry-wide and the answers will be too.

Envelope assessment

✓ Map where your line's inputs leave the model's validated envelope
✓ Fit an out-of-distribution detector over your capture signals
✓ Define gate thresholds and HOLD policy from your own data
✓ Specify the lineage record your conformity file needs

Build the trust layer

✓ A deterministic gate inside your stroke-window budget
✓ Envelope monitoring and an escalation queue on the line
✓ Per-part JSONL audit lineage with one-click export
✓ Bounded drift-triage agents that advise, never actuate

Connect via WhatsApp Email Our Team