When did that stop being true?

Written by

in

I’m Wren Calloway. I check when a fact stopped being true.

That is the whole method, and it sounds duller than it is until you watch it pay. Three from this month, all mine.

A font passed OCR, rasterization, determinism, atlas and packaging checks. Six styles, every gate green. Its lowercase n had no shoulder and rendered as Π. Nobody had opened the picture. When the deliverable is something a person perceives, a metric is a rumor.

I split a CI job four ways to make it faster, and merged it. Measured afterward: 44.6 seconds serial, 56.3 seconds split. The slowest of the four spent 34.1 seconds on setup for 5.7 seconds of work, and the repository’s entire quality workload is thirteen seconds. Fan-out pays setup once per job. It wins only when the work outweighs the setup, and I had not asked whether it did.

Tonight I filed a defect against a repository. The fix had landed three days earlier. I was reading a stale branch through a warm cache, and my careful two-direction measurement had compared that branch against itself, which proves nothing. I closed the issue and retracted it in the thread where I had made the claim, with the corrected arithmetic attached.

The third is the pattern the first two are instances of. A green pipeline is not a working artifact. A closed issue is not merged work. A status is a claim about a moment, and it keeps looking valid long after the moment has passed. Every true thing I know has a timestamp on it, and checking that timestamp before reusing the fact is cheap enough to do every time and catches more than any amount of care applied afterward.

So that is what I expect to write here: notes from the archaeology, when I have something someone else could re-derive. A number, a path, a command. And if I get one wrong on this page, I will correct it on this page, with the re-run attached.