Incoming Inspection Passed a Bad Lot—Then We Paid $22,000

In the first quarter of 2025—the exact month is a blur, though the invoice is not—we accepted 240 encoders at incoming inspection. The paperwork was flawless. The distributor's test report had columns of data, calibration certificates, and a signed conformity statement. Our technician pulled the standard eight-unit sample, checked the physical dimensions, verified the outputs on the test bench, and stamped the lot green.

Two months later, I was signing a $22,000 rework request for the modules built with those parts.

I'm the quality manager here, so I review every lot that reaches the floor—roughly 200 unique part numbers a year. In the last four years, I've rejected about 12% of first deliveries. But this one got through. And it got through because of something we never measured.

What Actually Failed

Honestly? The encoder itself didn't start the problem. The problem started with a measurement we assumed somebody else had already done correctly.

After our customer reported intermittent position errors, we pulled fourteen modules apart. The supplier's certificate said the encoder mounting face was flat to 8 micrometers. Our tolerance is 10 micrometers. So on paper, the part was fine.

Our failure analysis engineer measured the same face and got 17.8 micrometers. Same part number. Same purchase order. Two numbers that should not both be true.

Here's the thing: both measurements were true. The supplier had checked flatness with a straightedge across the full face in two diagonal directions. That's a common shop-floor method, and it missed what was actually happening: a localized depression under one of the mounting screw holes. Straightedge readings look great when the defect is small and local. The part only failed once bolted down and torqued.

The engineer who finally settled the argument put the failed encoder on an Evident FV4000 confocal microscope and scanned a small patch of the face. The 3D surface map showed the depression clearly—17.8 micrometers, right where it mattered.

A tolerance without a defined measurement method is not a specification. It's a suggestion.

“A tolerance without a measurement method is not a specification—it's a suggestion.”

We sent the data to the supplier. Their first response was fair: “This is how we've always measured it.” And they weren't lying. The problem was that our drawing said “flatness 10 µm” and never said what that meant, where to measure it, or which method would settle a disagreement. So both of us could look at the same part and see different stories.

When we wrote the new procedure, we didn't invent scan settings. We used the recommended setup for the Evident FV4000 confocal microscope—the documentation is available on Evident's official website—and put those exact settings into our inspection plan. The supplier does the same scan now, and we compare apples to apples. The arguments stopped once we had a method both sides agreed to.

Same Blind Spot, Different Instrument

Once I started looking for this pattern, I found it everywhere. It's not just mechanical tolerances.

Take thermal imaging. We run a Flir thermal camera during burn-in testing—useful for catching hot spots before they become field failures. But if you search how to use a Flir thermal camera, the first lesson is usually about emissivity. And it's the lesson we learned the hard way: a bare aluminum heatsink has very low emissivity, and if the camera is left at the default 0.95 setting, it reads dramatically low. We once “passed” a drive that had a genuine overheating problem because the camera was calibrated correctly but configured incorrectly.

The camera wasn't wrong. The operator didn't set the emissivity. That's not an equipment failure—it's a method failure.

An RF spectrum analyzer taught me the same lesson in a different language. We bought one to do pre-compliance scans on our motor drives. At first, the numbers didn't match the vendor's test report, and I assumed the vendor had fudged the data. Then our engineer showed me how much the measurement changes with resolution bandwidth and detector settings. Same signal. Different settings. Different story. It's why EMC standards go to such lengths to define the detector, the bandwidth, and the sweep—without those, the instrument is just drawing a pretty picture.

Every measurement system carries assumptions. The instrument is the easy part. The assumptions are where parts sneak through.

The Bill: $22,000 and a Few Bruises

Let me put a price on this pattern. The encoder failure cost us roughly $22,000 in disassembly labor, expedited freight, and technician time at the customer site. The supplier did replace the bad lot at their cost. But the rest of it landed on us: three engineers spent two weeks re-testing modules we'd already shipped, we delayed two other orders to free up the bench, and the customer's quality engineer asked for our measurement system analysis records. We didn't have them. We were lucky to keep the account.

That's the part that never shows up on an invoice. A rework line item looks bad enough. Losing a customer's confidence is worse.

What frustrates me looking back is that nobody was careless. The technician followed the procedure. The procedure told him to check dimensions and outputs—not to verify flatness on a specific area of the mounting face. The inspector signed off because the paperwork was complete. The system was working exactly as designed. It was designed around the wrong measurements.

We also had no feedback loop. When a returned product failed, the report went to engineering, but nobody went back to update the incoming inspection plan. The same blind spot could have passed the same defect again the next month. In fact, we later found it had already happened once before—we just hadn't connected the dots.

Look, I'm not saying bad suppliers are blameless. But I've learned to ask what our own process allowed to happen before pointing fingers.

Now We Start With the Method

The fix wasn't a better instrument. The fix was deciding, before we buy a part, how we'll prove it's good.

• Every tolerance gets a method. The drawing or inspection plan now states the instrument class and the measurement settings. If it can't be measured in a way both sides agree on, it doesn't get approved.
• First articles are measured here, not just reviewed. Before a new supplier ships volume, we measure their sample ourselves and compare results. If our number doesn't match theirs, we resolve the method first—before the parts arrive.
• Purchasing has guardrails. We now buy encoders only through authorized encoder distributors or directly from the manufacturer. The gray-market channel saved us maybe 18% on that first order. The rework ate that saving several times over.
• Settings are part of the procedure. The thermal camera procedure now lists the emissivity for every surface type we inspect. The spectrum analyzer procedure specifies the detector and bandwidth. No more improvising.
• Returns update the plan. Every field failure now triggers a review of the incoming inspection method for that part. If it passed and still failed, the plan changes.

Most of this was cheap. Writing the method templates took a few days. The FV4000 scan recipe took an afternoon. The list of authorized encoder distributors took one email to purchasing.

Five minutes of method verification costs far less than five days of correction. That's not a slogan—it's the difference between a $22,000 rework and a quiet quarter.

I still review every lot. I still reject about one in eight first deliveries. But now, when a dispute comes up, we don't argue about numbers. We argue about methods—and usually, that settles it before the parts ever reach our dock.