2026-08-27

A rule you wrote is not a check you built

On day one we wrote down a rule for this system, in a numbered decision document, with reasoning:

Credentials never enter the corpus. Passwords, PINs, card numbers, keys. They

go in a password manager. This is not negotiable, because it is the only thing

that makes reaching the system through an ordinary chat app acceptable.

On day two, a wifi password was filed into the corpus. Name and value, on a page about internet contracts, whose summary then helpfully advertised that the page contained credentials.

Nobody overrode anything. Nobody forgot the rule. The rule was simply text in a document, and text in a document does not inspect incoming messages.

It happened again, faster

Three days in, we wrote a second rule, about the public journal this entry appears in:

Everything published uses a fictional household. Not anonymised. Invented.

Redaction means catching every value, every time, forever, and one miss is

permanent.

Then we wrote the first three entries using the real vehicle's plate, its chassis reference, its type approval number and its actual insurance premium. Published them. Noticed a day later.

Same day as the rule. Same author as the rule.

Why it keeps happening

Not ignorance, which is the comfortable explanation. In both cases the rule was fresh, written deliberately, and in front of us.

What actually happened is more mundane. The real values were what was to hand. Writing an example about a model misreading a licence plate is easier with the plate that was misread. Filing a page about internet contracts is easier when you paste the note you already keep, and that note has the wifi password in it, because of course it does. The friction of inventing a substitute is small but it is not zero, and it arrives at exactly the moment you are thinking about something else.

A rule asks you to pay that friction every time, forever, from memory. That is not a control. It is a hope.

What changed

Credentials are now stripped before anything is filed. Deterministically, by pattern, not by a model noticing. The value is never stored, never logged, never echoed back, and the reply says what was removed and where it should go instead.

It runs on the verbatim copy too, which matters: the system deliberately keeps the original text alongside whatever it extracted, so the first cleanup pass missed the password entirely. It was sitting in the quoted original, exactly where it had always been.

The journal has fixtures now, and the rule about them is enforced by there being nothing else available to write with.

The general form

Every reliable behaviour in this system turned out to have the same shape, and every unreliable one had the other:

• Decided in code, or offered as a choice. Works.
• Written in a document and left to be remembered. Fails, eventually, usually by the person who wrote it.

Which is uncomfortable, because writing the document feels like doing the work. It is not nothing: without the rule you would not know what the check should do. But the rule is the specification. The check is the control. Shipping the first and calling it done is how you end up publishing your own licence plate three days after writing a document about never publishing your licence plate.