The number on our own landing page that nothing could check
Our marketing site reads its counts from the running product on purpose. The figures it did not read had drifted: a contract value out by nearly a million, an identity count off by one, and a governance claim in the Terms of Service that the code had stopped honouring.
This site has a rule about itself: every count on it is read from the thing it describes rather than typed into the copy. Route totals, the event catalogue size, notification counts. The reasoning is that prose is not tested, so a figure typed into a sentence drifts silently and nobody finds out until a customer does.
The rule was right and the coverage was not. An audit of every public page against the code found four figures that had drifted, and the worst of them was in a contract.
The demonstration project was advertised as a live £17.6M job, in a button, twice on the landing page and once more on the exposure calculator. The seed builds it at £18.5M. A visitor pressing "walk a live £17.6M job" arrived on a project showing a different number — which is the shortest route there is from marketing copy to a reader who stops believing the rest of the page.
The same page introduced its console panel as "drawn to the shape of the real seeded project rather than an invented dashboard with rounder numbers", and then carried a forecast margin, a delay exposure and four progress bars that appear nowhere in the seed. Below it, a payment scenario — application 14 at £1.42M against a valuation of £1.19M — was stated to run on the demonstration project. There is no application 14. The seed builds three cycles, and the third is a genuinely better story: applied £2,248,650, certified £2,129,000, £119,650 withheld for handrail terminations not to detail and dewatering rates not agreed, with a valid pay less notice served inside the window. It is more convincing than the invented one precisely because you can open it and read it.
The demonstration page said twelve identities beside a list of thirteen. That one had a date: a Construction Manager was added to the seed months ago and the sentence beside the list was not.
The serious one was not a number. Four pages and the Terms of Service said no AI agent in the system holds a mandate above PROPOSE. That stopped being true when two of the eighty-one agents were given an ACT ceiling — one that files a tender return register off a reading the platform already holds, one that says the platform itself is unwell. Both were deliberate, both are documented, and neither can reach a governance event, because the catalogue refuses AI authorship on those outright. But "no agent holds a mandate above PROPOSE" in a document a customer contracts on is a different kind of wrong from a stale figure on a marketing page.
What the true version says is stronger, which is usually how this goes. Almost the whole fleet can only propose. The two that may act do so only inside an envelope naming the exact commands, granted by a person holding governance authority, carrying an end date, revocable, recorded on the chain — and both carry a value ceiling of zero. That is a better sentence than the one it replaces, and it has the advantage of being true.
The fix in every case was the same shape as the original rule: read the figure from the thing it describes. The contract value comes off the seed through the platform's own money formatter. The identity count is the list's own length. The agent counts are computed from the registry, so the day somebody grants a third ACT ceiling the page says three. And the payment figures are pinned by a test that seeds the demonstration project and asserts the page and the record agree.
The lesson is not that we were careless with the copy. It is that a page which derives nine of its figures and types four of them will drift in exactly those four, and nobody will look, because the page has a reputation for being derived.
See the record being made rather than described.
Launch the demo environment