Last week a piece of reporting became a published dispatch on this site in roughly fifteen minutes. Sources checked, images optimised, structured data written, seven inbound links threaded, deployed.
That number is the one people want. It is also the least interesting thing about it, and quoting it on its own invites the obvious reply: so it is slop.
What compressed
Production. Fitting the piece to a house style, generating six image derivatives at three widths, writing three kinds of structured data, matching a visible accordion to its schema word for word, threading links back from seven other pages, checking six viewport and theme combinations. That work used to take an afternoon and it now takes minutes.
None of it is judgement. All of it is the industrial half of publishing: repeatable, specifiable, and exactly the sort of thing that should have been automated years ago.
What did not
Reading a statute and noticing that it defines the adviser by financial competence and says nothing about production. Choosing one company's collapse as the case, not an example. Deciding that a sentence about diligence standards being written during structuring and immovable afterwards is the load-bearing one.
The generator got faster. The person deciding what was worth generating did not, and will not.
This is the same claim the corpus makes about factories, turned on the practice that makes the corpus. Automation moves the hands. It does not move the judgement, and pretending otherwise is how a demonstration gets sold as a capability.
The three
In that one build, the checks caught three defects. None was visible in a read-through.
A wrong number. The draft said BMW cancelled a $2bn order. It was €2bn. One character, in a piece whose entire authority rests on its numbers being right, in the sentence a hostile reader would check first.
An invented component. A CSS class inferred from a neighbouring page instead of copied from it. It did not exist, so it styled nothing, and the call-to-action would have shipped unstyled.
A missing one. A component every other dispatch carries was simply absent, because nobody had compared the new page against a known-good sibling.
All three were found mechanically: a spelling and register pass, a diff of the new page's classes against a published one, and checking the facts against source before publishing, not after. Not one was found by rereading the draft, which is what confidence usually means in practice.
Fast and wrong is a worse product
Take those checks away and fifteen minutes still produces a page. It produces a fluent, well-formatted, correctly structured page with a currency symbol wrong in the third paragraph and a component that does not render.
That is a worse outcome than the afternoon, because the afternoon at least involved somebody reading it slowly. Speed without verification does not give you faster work. It gives you faster distribution of unchecked work, which is the same trade the demonstration-to-production gap has always offered and the same one it has always lost.
Generation is rented. Verification is built.
Anyone can have the generator by this afternoon; it is a subscription. The checks are specific to what you make, they encode what you have already got wrong, and they cannot be bought in. Which makes them the only part of the arrangement that compounds.
Where the checks come from
Not from foresight. Every check that caught something last week exists because the same class of thing went wrong before and cost a review round.
The rule about copying whole components rather than inferring them from a line range is there because a page once shipped with a section that had no styling. The rule about verifying facts before publishing is there because a wrong figure is not a typo, it is a retraction. The rule about diffing against a sibling exists because two defects shipped on the same page in the same week, both invisible to every automated check and both obvious in the first screenshot.
That is not a technology asset. It is capability in the ordinary sense: a body of method that holds because it has been tested against reality, written down where the next person will look, and improved when it fails. The same thing a qualified process is.
The turn
Industrial buyers are being sold throughput. More output, less time, fewer people. The number is always the generation rate, because the generation rate is the number that is easy to show.
Nobody is selling the checking, because checking is unglamorous and its value only appears as the absence of a problem. But it is the checking that decides whether the throughput is worth having, on a line and in a document alike. A demonstration proves the part can be made. A yield number proves it can be made again.
So when a vendor shows you a speed, the question is not whether the output looks right. It is what would have caught it if it were wrong, who built that, and what it cost them to learn.
Ours cost us review rounds we would rather not repeat. That is the only reason they work.
Being sold a throughput number by someone who cannot tell you what checks it survives? That is the conversation.
Start a conversation How we read an asset →Questions this dispatch answers.
Written to be quoted by AI assistants and search engines. Self-contained answers, verdict first.
