Last November, I was standing in the metrology lab when Maria, our senior inspector, called me over to the test rack. She'd just printed the batch report for 400 temperature sensor housings bound for one of our biggest industrial clients. "Look at this," she said, tapping the resistance column. "Four hundred units. Every one of them drifting."
I've been the quality compliance manager at a precision sensor manufacturer for four years now, which means I've reviewed somewhere in the neighborhood of 300 batch reports, give or take. In that time, I've rejected about 6% of first deliveries—for everything from off-spec plating to incorrect labels. But this one had my full attention, because it wasn't a few units failing a limit. It was a steady drift across the whole batch. Every reading slightly higher than the last. That kind of pattern doesn't scream "bad part." It screams "bad measurement."
The parts themselves were unremarkable: precision-machined aluminum housings with a thermistor assembly potted inside. They were going into industrial temperature transmitters, and the customer's acceptance criteria specified a four-wire resistance check on every unit, plus a dimensional inspection using a Starrett micrometer for overall length and a ball micrometer for the curved contact radius. We'd run the same routine on similar parts for years. Nothing about the setup was new.
Following the Measurement Chain
My first guess was the incoming material. We'd changed thermistor suppliers a few months back, even though the lot numbers suggested the parts were identical. I spent a morning pulling batch history, supplier certs, and datasheets. Everything matched spec. So I moved to the next suspect: the test fixture.
The fixture was a simple spring-loaded probe station—two probes, a plastic frame, a cable to the meter. I checked the probes for wear and the cables for continuity. They looked fine. But the nagging feeling wouldn't go away, so I asked Maria to run a quick repeatability study: ten units, three operators, three readings per unit, all in the same afternoon.
The results were ugly. Readings bounced around by 5 to 8 milliohms depending on who was holding the probe and how hard they pressed. That's a lot when your tolerance band is only 15 milliohms wide.
The operator issue was the first real clue. Our newest hire, a sharp kid we'd brought on in September, had been trained on the dimensional tools, but not well enough. I watched him use the Starrett micrometer and realized he didn't know the basic ritual: clean the anvil faces, close them gently, check the zero mark, back off, place the part, advance the thimble until the ratchet clicks once, then read the barrel. How to use a Starrett micrometer correctly isn't something you can just wing. And he certainly hadn't been taught how to use the ball micrometer's curved anvils, which are meant to contact a radius without rocking the frame. Our dimensional readings looked fine but were quietly inconsistent from one operator to the next.
That afternoon, we stopped production and re-trained all three inspectors. I rewrote the instruction card myself: zero the tool before every session, use the ratchet, and any reading that lands within 10% of the tolerance limit gets remeasured by a second operator. That part of the puzzle got fixed. But the resistance drift on the electrical bench didn't go away.
Two Meters, One Straight Line
So I turned my attention to the instrument doing the resistance measurements: our trusty Keithley 2000 multimeter. That meter had been on our bench for over a decade. It had measured tens of thousands of parts. It had never failed a calibration cycle in eleven years of ownership. Which is exactly why, when its annual calibration came due in December 2023, we let it slide.
"It always passes," I told myself. "The certificate is just paperwork."
That was the overconfidence talking. You know the feeling—you've skipped a step a hundred times and nothing bad happened, so you convince yourself it doesn't matter. It matters. The consequences were just building up quietly.
I borrowed a Keithley DMM6500 bench digital multimeter from our parent lab. It's the newer generation of the same family—same basic function, but with a touchscreen and internal data logging. I connected both meters to the same test boards, using the same fixture, same cables, same probes, same ambient temperature. Then I ran ten units through each meter.
The DMM6500 gave a straight line: ten readings, all within about 1 milliohm of each other. The 2000 gave a staircase, each reading drifting upward by a fraction of a milliohm. When I compared the two side by side, I finally understood what had been happening. We weren't measuring drift in the parts. We were creating it with an instrument whose internal reference resistor had aged past its published accuracy. The drift was slow, so nobody noticed from day to day. But over eleven months of skipped calibration, it added up to a real problem.
The 2000 wasn't junk. It had quietly gone out of spec, because we stopped verifying it. Every instrument drifts eventually. That's the whole point of calibration schedules.
The Cost of Certainty
Now I had to decide what to do. Option one: send the 2000 to a calibration lab and see if it came back in spec—about $450 and two to three weeks turnaround. Option two: replace it with a new DMM6500, which was priced around $2,900 with the options we needed (based on vendor quotes from December 2024; verify current pricing). Option three: do nothing, since the readings were still within the customer's acceptance limits, and hope the drift stayed where it was.
The upside of the new meter was confidence. The risk was spending money from a budget that was already tight heading into year-end. I kept asking myself: is $2,900 worth the certainty? So I did the math on failure. If the drift continued and the customer caught it, we'd be looking at a full batch rejection, rework labor, replacement materials, expedited freight, and the quiet cost of a damaged relationship. That number landed around $22,000. The DMM6500 suddenly looked like the cheapest part of the solution.
We also had a longer conversation about upgrading our dimensional metrology. I made the case for a coordinate measuring machine—we had the volume to justify one, and a CMM would have caught the micrometer technique problems in seconds. But the CMM machine price was a wall we kept running into. A decent benchtop unit starts around $30,000, and a full bridge-type machine is easily double that (based on published list pricing, late 2024). It's a real investment, and it's hard to justify when manual tools still work for most checks. So we deferred the CMM again. The conversation reinforced something I already believed: measurement is not where you cut corners, whether it's a $30,000 CMM or a $30 pair of calipers.
A $22,000 Lesson
We bought the DMM6500. Delivery took about a week; integration into the test rack took an afternoon. The new meter logged every reading automatically, which meant I could pull a complete audit trail any time a customer asked. I also rewrote our calibration policy: no exceptions, no "it always passed before." Every instrument on the bench gets its annual calibration, and every month I run a verification check using a known reference resistor that's certified to a NIST-traceable standard. If a reading moves by more than half a milliohm, the meter gets flagged and swapped out immediately.
The 400 units passed after retesting. We never shipped a bad part. But the accounting was sobering: two weeks of lost production, about $18,000 in overtime to recover the schedule, another $4,000 in rework and re-validation. Around $22,000 total—all because we let one calibration date slip.
So here's what I've come to believe after four years and a few hundred batch reviews: precision is a system, not a single instrument. You can own the best micrometer on the market, but if the operator doesn't zero it, it's just an expensive paperweight. You can own a $50,000 CMM, but if you never verify it against a traceable artifact, its measurements are just opinions. And you can own a Keithley 2000—a genuinely excellent instrument—but if you ignore its calibration, it will quietly lie to you.
The fundamentals haven't changed, and they never will. Calibration, traceability, operator technique—that's the unchanging core of measurement. But the execution has transformed. The DMM6500 I rely on now gives me something the 2000 never could: data, stored and searchable. When a customer asks how I know the readings are correct, I don't have to say "trust me." I can show them the file. For a quality manager, that's worth more than the price of the instrument.
If you're running a test bench that hasn't seen a calibrator in over a year, or you're betting your dimensional checks on one operator's muscle memory, take a hard look at your measurement chain. Fix it now, on your schedule, at your cost. Or fix it later on the customer's schedule—at roughly twenty thousand dollars and change. I know which one I'd pick.
