The machines I never tested

For most of the summer, laptop and integrated graphics were handed the middle version of a capture rather than the sharpest one. The code said those machines could not be trusted with the top file, and I had stopped reading that line some weeks earlier because it looked settled.

So I went to find where it came from. It came from May. From a card that could not build the largest scene we had without the browser dropping its graphics device entirely, which is a black screen and a reload. That happened, it was real, and writing the rule was the right call at the time.

It was a measurement about a car driving through a different scene, and it had never been taken again: not on foot, not on the smaller town captures, and not after two months of work on how much of a place is held in memory at once. A fact about one afternoon had been stored as a fact about a class of computers.

This is the failure I am most prone to and it has nothing to do with skill. A number gets measured honestly, gets written into a rule, and the rule quietly outlives the conditions that produced it. Nothing fails. Nobody sees a bug. The software is just a little worse than it needs to be for everyone in a category, forever, on the strength of a Tuesday.

What replaced it: every machine is offered the sharpest version, and the frame rate is what argues it down. That is a policy rather than a guess, and the thing it depends on is re-measured every session by the machine it is about.

The old approach also had no way back. If a machine genuinely could not hold what it was offered, the browser dropped the scene and the page had nothing left in it to remember why, so the next visit walked into the same wall at the same speed. That failure now leaves a note that survives the reload, per place. It expires after a month, because a bad afternoon is not a permanent verdict about somebody's computer, which is the mistake I had just spent a week undoing at a larger scale.

I have become wary of writing rules of this kind at all. The ones that survive contact with a year seem to be the ones that measure something rather than classify somebody.