We can't see each token on its way out, but if I play this out in my head, I can see it generating ['</', 'output', ...] and jumping to '_list'.

Turns out that to an LLM, it tokenizes

as ['']

which looks too much like the tokens for closing ['']

You can see this in - platform.openai.com/tokenizer

It mixes up the two adjacent tag closings into one.

Changing the tag names fixes it

Also the drafts and the entire blog text is actually maintained in a Subversion repo, which I keep planning to migrate to git but never have.

The txt -> html is driven by a Makefile pattern which now works on BSD make (no GNU make magic).

Fixing this is a whole yak shaving exercise.

Probably the tap cartridge in the bathroom which gets slapped down by kids.