The Pages That Still Had a Scout

For most of last year our pipeline ran as a set of named agents. A Scout that hunted for products. A Researcher that checked whether a brand could be trusted. A Judge that ran the scoring. A Sentinel that clicked every link before anything shipped.

They were real. They ran for hours on each product, and a large part of what is on this site today came out of them.

We changed that design. The work now runs as passes and laws, with one AI and a human gate at every step, and no characters. What we did not do is update the pages that mentioned the old design in passing.

What was actually wrong

Not the history. The tense.

Several live notes described the retired process in the present tense, as though it were still how the site works. One named the internal version of a discovery agent that checks a table before every piece of research, which stopped being true when that agent stopped existing. Another said our agents now hold institutional memory.

A reader landing on those pages would form an accurate picture of a system we no longer run.

One page also carried a real error inside its own record. It described a run in its body and summarised that same run in its summary field, and the two named different numbers of agents: seventeen in one place, sixteen in the other. Two counts of the same thing on one page, published for months, because nothing ever compared them. Seventeen is the right number. Both now say so.

What we deliberately did not change

Everything that was a record of something that happened.

Where a note says a Scout found a soap in January, a Scout found a soap in January. Where a transcript shows an agent refusing to start until a human gave it two missing inputs, that is what the machine printed on the day. We could have swept those names out and made every page agree with today’s vocabulary. That would have been tidier, and it would have been a lie about our own history.

So those pages keep their accounts intact and now carry a dated note saying which generation of the system produced them. A reader can tell history from present tense, which is the only thing that was ever broken.

What did not stay

Two pages came down on the same pass, and the line between them and everything above is worth stating plainly.

Both were whole articles whose entire subject was the old design, written in the present tense as though it were how the site works now. One of them still headed a section with a count of checkpoints that no longer exist. Neither was an account of anything that happened, so there was no record inside them to protect, and keeping them up would have meant publishing a description of a system we retired. Both addresses now lead to a page that explains how the site actually works.

What stays is what happened. What went is what claimed to be current and was not.

This section was added on 8 August 2026. The rule above was published on 7 August and did not say where its edge was. This is that edge, written down.

Why it took this long to notice

Because nothing was looking. Deleting the pages that described the old design felt like finishing the job, and by then the old vocabulary had spread into places nobody thinks of as pages: a caption under a chart, a summary field, one clause inside a note about something else.

There is now a check that reads every piece of published content we hold and marks any sentence still describing the retired design. It prints how much it actually read before it reports anything, so a check that quietly scanned nothing can never come back looking clean. It runs on every new piece we build.

It also found something we have not fixed yet: an invisible attribution sitting in the source code of dozens of notes, crediting a chart to an agent by number. That one is logged, in the open, and still open.