Copying became free. Checking did not.

This week's case is called ArtCraft. The founder of a synthetic-voice company, working with a coding assistant, published twelve programs in twelve days: copies of Photoshop, Illustrator, Premiere, Acrobat and, since October 7, Word, Excel and PowerPoint. The Photoshop copy had 806 thousand downloads in ten days. The press called them "clones".

We downloaded the repositories and read what they publish themselves. 449 thousand lines of code, 1,494 changes in twelve days, 1,291 of them co-signed with the assistant. And a scoreboard, generated by a program and checked on every change, that says the opposite of the headline: "ready for real work: 45 %". Speed: 3 of 56 measurements within target. Changing a layer's opacity takes 443 milliseconds where the target is 20. And 31 of the 154 settings do nothing: they are there to look right.

The most honest number on the scoreboard is the files. They made 258 files inside the real Photoshop, with the image Photoshop draws stored inside each one, and compare pixel by pixel every night. Today 52 % of those files open identically. The rest do not.

Notice the split. Photoshop files have an answer key: the original draws, the copy draws, you compare. That number was lower last week and rises every night, because every difference becomes a task with a right answer. Speed, the feel of use and the settings that do nothing have no answer key: nobody wrote "fast enough" in a form a machine can check. Those numbers do not move.

The same rule explains why a spreadsheet is easier to copy than a photo editor. Every spreadsheet calculation has a right answer you can write into a test. "This photo looks good" has none. It also explains a case from September: a programmer rebuilt a 1989 game in a week with four hours of attention, because the original game ran alongside, serving as proof for every screen.

The copy converges at the speed of the answer key. Without one, what you get is a façade that looks finished and is not, and nobody knows by how much. That is the difference between a measured 52 % and an estimated 45 %.

In the same project, the pipeline that checks every change cannot keep up: of the last hundred pushes to the main line, 83 were cancelled because the next one arrived before the check finished. Three hundred changes a day come in; the queue of checkers is longer than the queue of makers.

In June a group of mathematicians measured the same thing in money. They gave ten unpublished problems to four assistant set-ups and paid thirty specialists to grade blind. The bare assistant, with no apparatus around it, got 5 of 10 right for 117 dollars. The most expensive set-up got 6 of 9 for 3,186 dollars. The one with a Fields medallist on the team got the same 5 for 4,799. The expensive apparatus bought one extra answer at 27 times the price. What actually cost money was the thirty graders.

It is the arithmetic this notebook has repeated since "The cheap one acquits": when making tends to zero and checking does not, the checked fraction falls by itself. Asking for more care does not help. The scarce work became judging, and that is what gets charged for.

The large company does not pay for the editor. It pays for the file format legal accepts, the login HR controls, the support that picks up, and for not having to retrain two thousand people. Thirty dollars a month per seat is cheap next to that. So, for a year or two, the copies show up in nobody's revenue. They show up in renewal discounts and in law: New York's rule forcing one-click cancellation of subscriptions came into force this week.

What the copies destroy is the belief that the program itself is worth the subscription. And they expose what is: the file format, which is the answer key every copy is measured against; the model and the data the program wraps; and the receipt, whoever signs the contract and answers for the breach. No twelve-day copy carries those three things.

Who really bleeds is the niche program at 50 to 500 a month whose value was a screen over a database: customer lists, quote generators, PDF tools. Those have a trivial answer key, no format of their own and no auditor at the door, and the customer already pays for the assistant that rebuilds them in an afternoon. The copy aims at Adobe for marketing. The blood runs in small software.

The cloud splits in three, and the middle one shrinks. The model cloud stays: the inference machines nobody runs at home. The edge stays: address, tunnel, identity, defence, which one company is buying piece by piece. The application cloud shrinks, the rented machines running screens over databases, because that becomes a twelve-megabyte file running on your own machine for free. This same week, three data centres of a Russian company were hit by drones in four days; redundancy within one country stopped being redundancy.

Running on your machine comes back through the arithmetic, not the ideology. When generating the program costs nothing, renting someone else's machine becomes the biggest cost and the biggest risk. A site published this week by a one-person consultancy runs on a home computer, with no public address and zero hosting cost, and can be audited end to end in an afternoon. In three years, "runs on your hardware" is a pricing tier, not an exception.

An open format with a public answer key becomes the move. If the moat is the format, the attack is an open format with reference files and reference images kept by a consortium, not by a clone. Whoever controls the answer key controls the market for copies. It is a place where a small team with measuring discipline is worth more than a lab with a thousand graphics cards.

"Who reviewed this?" becomes a purchasing question. Readers already abandon machine-written text when they notice. Software goes the same way: a published and verified scoreboard, a record of who looked, the name of who signs. ArtCraft did half of that by instinct. It will become a requirement in tenders and in insurance.