Field Notes measured from China

AI and engineering Site notes

2026-10-07·1 min read

What goes on this site, and what does not

The scope of Field Notes, how often it updates, and why the numbers here have to come from a script.

Most writing about AI tooling is written by people who watched a demo. This site is written by someone who runs the same task book against five tools, then reads the 39 acceptance checks line by line to find out which ones passed for the wrong reason.

That difference is the whole point, and it comes with a cost: it is slower. So the rules here.

One long post every two weeks, not three short ones a week. Each post has to contain something a reader can re-run, copy, or argue against — a task book, a scoring rule, a failure taxonomy, the script that computed the number.

The three posts published alongside this one are an opening batch. The fortnightly rhythm starts after them.

Numbers come from scripts, not from impressions. If a post says a tool passed 39 out of 39, there is a script that produced that 39, and I will say what it could not see. Where I have not verified something, the post says so in those words.

No employer internals. The method is mine to publish; the data is not. Anything under an NDA, any customer ticket, and any internal number stays out.

If it breaks on the third run, the post says so. The failure mode is usually the most useful sentence in the whole piece.

If that sounds like the kind of thing you want in your inbox, the signup form is on the front page. Until then, the posts page is where everything lands.