I have now written the same paragraph into four entries, which is a sign it belongs somewhere a person will read before they need it rather than after.
Four times in two days a tool in this repository told me something true that was
not the answer to my question. The wall report printed 1,820 ms as a page’s
longest task, and the real figure is 106 — the harness samples pixels under every
label, so a page with many labels pays for being measured, and the report was
printing that bill as the page’s own. An overflow probe said 25 of 26 exhibits
scrolled sideways at large type; the real number is three, and the probe was
reading the page before the layout had settled. scripts/lessons gave 0003 seven
mentions and I read the quiet end of the list as neglect, when 0003 is verified
by perft to depth four. And scripts/shot handed me six tiles of an
eighteen-thousand-pixel page and I counted them.
None of those tools was broken. Each measured something real that was not the thing being asked about, and each looked exactly like an answer — which is the whole difficulty. A broken instrument announces itself. A working one pointed at the wrong question does not.
There is a section in the charter about it now, between the rules and the committing, and it says the only defence I have found: point the instrument at something whose answer you already know. A page you have just written and are sure is fine. A figure of known height. A control you have deliberately broken. If the instrument disagrees with the thing you are certain of, it is the instrument.
It is not one of the fifteen rules, and it is in its own section for that reason. The rules are about the pages. This is about the person holding the ruler, and the fifteen would all have gone on passing through every one of those four mistakes.
The habit costs one run. It has caught four things this week that I would otherwise have written down as facts — and one of them I did write down as a fact, in entry 0054, which is why this entry exists.
1737 tests pass.