NVIDIA's technical blog carries a pitch quality engineers have heard before: AI agents that validate a factory change on a digital twin before the change ever reaches the line. The zero-scrap trial run. The approval that costs nothing in material, downtime or customer risk. I have spent two decades running the other kind of trial – real scrap, at rate, on the shift that actually generates your PPM. So here is an uncomfortable observation. The twin approves the process you assumed. The plant runs the one you have. Those are not the same process, and the distance between them is where your defect costs live.

The twin validates the process you assumed

Digital twins are built from nominal. Fresh tooling. Material at mid-specification. Cycle times lifted from the process sheet. Fixtures straight out of the CAD model, locator pins at zero cycles, weld tips dressed that morning. Point an agent at that and it will faithfully discover that a nominal process makes nominal parts. The agent is not wrong. It is answering a different question from the one your customer will ask.

Failure modes live in deviation. The locator pin at 40,000 cycles carrying 0.08 mm of wear. The lubricant viscosity that moves between lots because a supplier quietly changed a release agent and told no one. The torque wrench the night shift knocked off the bench at two in the morning and hung back on its hook. I have never seen a twin that models the wrench that gets dropped.

A QRQC case from my SNOP years makes the point. A bent-tube cell, chassis part, outside-radius thinning. First articles: clean. Capability study: clean. Prototype loop: clean. Then hour two at rate, and the rejects started. The mandrel had warmed and expanded by fractions of a millimetre, and the material stopped flowing where every simulation said it would flow forever. Containment took a shift. Scrap ran to five figures in euros. The defect did not exist at nominal, and the twin lived entirely at nominal.

The physics arrives late

Heat builds over an eight-hour run. Tool wear curves inflect only after tens of thousands of strokes. Fixtures fatigue. The operator at 05:00 seats the part differently than the process engineer did at buy-off. The first simulated hour always passes. Physics arrives late, and it arrives at rate.

Building the QA/QC function at the SNOP greenfield plant – 900+ people, launched from a bare shell – taught me the cleanest version of this law: everything is perfect at commissioning and everything is different at volume. The capability study on day one and the PPM reality of month four are separated by precisely the variables a twin abstracts away. We cut defect costs 70% on that ramp, and every point was earned on the line, not in a model.

This is why APQP exists. IATF 16949 expects a significant production run – real tooling, real operators, real material lots, at contracted rate – because the people who wrote the standard knew where defects hide. A run-at-rate is not ceremony. It is the only event in product development where nominal and deviation are locked in the same room.

A twin is a quality record – treat it like one

None of this makes twins useless, and NVIDIA's agents will earn their keep. Run the change in the twin first and you kill the stupid options for free – real money in a world where a trial run costs tooling, material and a shift of capacity. Agents are already doing design checks in semiconductors at TSMC; AI inspection is scaling at Siemens and Procter & Gamble. The pressure on quality leaders will be to accept simulation evidence in place of the trial. Some of that pressure will be justified. The terms matter.

If twin evidence substitutes for a run-at-rate, the twin stops being an engineering convenience and becomes a quality record. That brings obligations. Calibration against live line data, on a schedule. Version control, so you know which twin approved which change. Revalidation triggers when the process drifts, the same way you requalify a CMM. Leave the twin outside your control plan and audit scope, and you have a PPAP on file for a process that no longer exists – the first thing a good customer auditor goes looking for.

I build multi-agent systems myself. MultiPS, my orchestration platform, runs 63+ models in parallel with consensus synthesis, and it taught me a specific humility: when the models agree, they have agreed about the question they were given. Sixty-three models reaching consensus inside a simulation have reached consensus about the simulation. Agreement without ground-truth line data is agreement about the model. A compass heading, not a destination.

Simulation is where assumptions go to be agreed with.

Key takeaways

  • Use the twin to narrow the search space before the trial – never to replace the run-at-rate that APQP and IATF 16949 exist to enforce.
  • Calibrate the twin against live line data on a schedule, and treat twin-to-line drift as a nonconformity, not a curiosity.
  • Version-control the twin and define revalidation triggers in the control plan if its evidence will justify process changes.
  • Model consensus without line data is consensus about the model. The line arbitrates; the agents only argue.

So take the zero-scrap trial run, but take it on these terms. Run the change in the twin, eliminate what the agents can eliminate, then prove the survivor at rate on the shift that generates your PPM – with the tooling, the material and the people you actually have. Feed the twin line data until it converges on the plant you operate rather than the one on the process sheet. The line is the only ground truth in the arrangement. Everything else, however many agents vouch for it, is a model wearing a lab coat.