How a product earns confidence
One pass is not enough. Confidence needs enough samples.
How much evidence is enough?
Every NPI build ends with someone deciding to move on. My work is the evidence behind that decision, and the hard part is knowing when there is enough of it.
Statistics answers most of that question. Not all of it.
Three passes are not proof
Three units pass a test. It feels like good news. Statistically, it says very little.
To claim 90% reliability at 90% confidence, 22 units have to survive the test with zero failures. Three passes support a claim of about 46% at the same confidence.
The count only matters if the measurement can be trusted, which is why Gage R&R comes first. Sample size limits what a result can tell you.
A good average can hide early failures
Failures do not arrive evenly. Over a product's life, the failure rate often follows a bathtub curve: early failures from defects that escaped the line, a flat stretch of random failures, then wear-out.
Each region needs different evidence. Screening and build quality catch early failures. Life tests show wear-out. A model that assumes a constant failure rate, like the exponential, can miss both ends.
Average lifetime alone can be misleading. Take a synthetic example: 70 units cycled to failure. Three models fit the data with nearly the same MTTF, about 22,000 cycles. Their predictions for when the first 10% fail differ by about four times. The fitted Weibull shape, β ≈ 2.2, says the failure rate rises with use. The exponential model assumes it stays flat, so it fits worst.
Warranty returns live in that early tail, and the tail depends on the model. A mean that clears the target can still hide a problem. Until the data can tell the models apart, there is not enough evidence yet.
A failure is not yet an explanation
Statistics can tell you that something is wrong. It cannot tell you why. That is the work of failure analysis and corrective action (FACA).
Root cause is the chain of conditions that made the failure possible. The failed component may be the endpoint, not the origin. Design, process, material, firmware and use environment can all belong to the same mechanism.
A corrective action matters only when it breaks that chain. Revalidation shows whether it did.
Passed tests are a start. Confidence comes from knowing why things fail.