Zero defects is the goal. Zero evidence is the trap.

Here is an uncomfortable observation. The most dangerous dashboard I ever stood in front of was the quiet one. Flat green line where the red spikes used to live, in a plant that had just gone six weeks without a recorded nonconformance. Everyone around me wanted to celebrate. I wanted to know what had broken. Across two decades in automotive and aerospace I have learned that silence in a quality system is a claim, not a fact. So when Alstom's Bhavik Pathak describes AI predicting defects before they occur, and Rockwell wires FactoryTalk VisionAI straight into Plex QMS so that vision systems and quality records share one nervous system, I see real progress – and a trap nobody has priced. A control chart with no points is not a victory. It is an uncalibrated instrument.

Quality systems run on failure

Strip the labels off QRQC, 8D, A3 and PFMEA and you find one machine underneath: nonconformance in, knowledge out. At SNOP we cut defect costs by 70%, and the mechanism was embarrassingly unglamorous. QRQC forced every defect – line stoppage, customer complaint, a cracked weld caught in an audit – into a structured reaction within 24 hours, treated as data rather than paperwork. The reduction came from the volume of failures we honestly captured, grouped and counter-measured. Record less, improve less. That is the whole engine.

The same plant was a greenfield, and I built its QA/QC department from bare concrete for more than 900 people. Designing feedback loops from scratch, instead of inheriting them, teaches you something unforgiving: loops need fuel. A PFMEA stays alive only as long as failures keep editing it. An 8D library with no new entries is a museum. Starve the failure stream and your risk knowledge freezes at launch – you run the plant on launch-day assumptions with launch-day confidence. A predictive system that succeeds does exactly this, by design. It removes the signal that both its learning and yours depend on.

The silence problem

Zero is the least informative number in quality. Three failure modes hide inside it.

• Escapes you cannot measure. When complaints stop, you cannot distinguish a clean process from a customer who has quietly added sorting at their receiving dock and stopped telling you. Silence downstream is ambiguous evidence.
• Drift you cannot see. The predictor was trained on yesterday's defect distribution. Suppliers requalify materials, tooling wears, a seasonal humidity shift changes a curing process – the input distribution moves and nothing notifies you. The model's silence means less every month, and no alarm is attached.
• Statistics that refuse to cooperate. With zero defects in n units, the rule of three puts the 95% upper bound on the true defect rate at roughly 3/n. Ten thousand consecutive good parts still permit a true rate near 300 ppm. In automotive volumes that is a field campaign, not a rounding error.

This is where the industry pitch overreaches. The promise is not merely fewer defects; it is that inspection ends. Inspection does not end. It gets promoted – from auditing parts to auditing the predictor – and the promotion comes with a salary almost nobody has budgeted.

Zero defects is a result. Zero evidence is a belief with a dashboard.

Manufacture your own evidence

If success deletes your ground truth, you manufacture it back. None of this is exotic.

Seeded challenge lots first. Quarantined batches carrying known, deliberately planted defects, pushed through on an unpredictable schedule. I hold an ethical hacking certification alongside the quality ones; this is penetration testing with part numbers – probe your own defences before the field does it for you. A system that calls a seeded lot 100% good has not passed. It has failed in the most informative way available to you.

Escape audits supply the ground truth. Audit finished goods at the customer boundary with a hostile customer's mindset, on a fixed cadence, and treat that result as the reference the model is judged against. The audit does not shrink because the model is confident. The audit is why you are allowed to be confident.

The model is a gauge, so give it gauge treatment. Repeatability: same part, ten runs. Reproducibility: across shifts, lighting, operators. Bias: measured against the escape audit. Chart the detection rate on seeded lots as a live KPI, and notice the recursion – the day that chart goes quiet, you have the silence problem again, one level up.

Running MultiPS – an orchestration platform I built that runs 63+ AI models in parallel with consensus synthesis – taught me the hardest version of this lesson: consensus without ground truth is just agreement. When every model nods at once, I have learned something about the models and nothing about reality. The shop-floor translation is exact. Three algorithms agreeing is not proof. Proof still has to touch a physical part occasionally, and someone has to fund the hand that touches it.

The economics are not close. A seeded-lot programme with quarterly escape audits costs a few thousand euros a year. One containment and re-sort triggered by a single genuine escape across a European customer base starts around €50,000 and rarely stops there.

A predictor that never sees another defect is not a success story. It is an uncalibrated gauge with good PR. The plants that come through this decade intact will be the ones that kept paying for proof after the dashboards went quiet. Prediction does not kill inspection; it promotes it, from checking parts to checking the thing that checks the parts. That is the cheapest permanent insurance a quality organisation will ever buy.