A customer rejected a batch in March on the strength of his own incoming test, which disagreed with our certificate by about six percent. We retested, sensibly enough, by splitting a sample between our two laboratories. They came back four percent apart from each other.
Nothing was faulty. One laboratory uses an instrument bought in 2011 and the other one from 2021. Sample preparation is written down at each site separately and the two documents differ in two steps. One reports to two decimal places and the other to whole numbers. One records at ambient temperature and the other corrects to twenty degrees. And the field in our laboratory system where the method goes is free text, which over nine years has collected twenty seven spellings of four procedures.
The expensive consequence was not the batch. It was a trend analysis run last year across three years of results, which told the product team that a formulation had drifted and started a piece of reformulation work. It had not drifted. After a reorganisation in 2024 the share of samples going to the newer laboratory rose from a fifth to roughly two thirds, and what the chart showed was a change in routing.
Our system stores a result and a date. Method, instrument, operator and procedure revision were either absent or typed in by hand. Nobody ever specified them, and I understand why: to the technician the method is not information, it is the bench she is standing at.
Method is now a controlled list with a revision number, instrument is an asset reference, and results are held at full precision with the reported rounding stored separately. Any report crossing a method revision or a site says so at the top instead of averaging quietly. The two sets of procedures were reconciled by a chemist over five weeks, and neither turned out to be wrong, only different. Each quarter both laboratories test a split sample and the difference between them is recorded, which gives us a number to argue with rather than an assumption to rely on.
We spent nine years building a database of results and believed we were building one of measurements. A figure without the method that produced it is not comparable with the laboratory next door, and not with itself two years ago either.
– Serguey Shinder
