The open web is filling with text that no person wrote and no person checked. It is cheap to make and it reads well enough to pass. Every month there is more of it, and every month the average page is a little further from anyone who did the work it describes.

Models trained on that output drift. When a system learns from its own reflections, the rare cases go first, then the specifics, and what is left is a confident average. The finding has a name and a citation: Shumailov and colleagues published it in Nature in 2024. The practical version is simpler. Copies of copies fade.

The supply of verifiably human text from before 2022 does not grow. It is a fixed pool, and it is being spent. So the interesting question is not how to find more clean data on the web. There is no more. The question is where clean data still gets made.

It gets made inside companies. A claims desk, a loading dock, a returns counter, a night shift on a press line. People make thousands of judgment calls a day, and almost none of it is recorded in a form a model can learn from. The knowledge exists. The record does not.

That is the gap Ultro Labs works in. We run experiments inside companies to produce the record: video, audio, screen capture, sensor traces, and decision logs, all tied to the person who made the call and the reason they gave. It is slow, it is physical, and it cannot be scraped.

We ship every dataset with its own eval. A dataset without a test is a claim. A dataset with a test is an instrument. If a model is not measured on your own ground truth, any number attached to it is a guess about someone else's problem.

Data made this way behaves like capital, not like a subscription. It does not expire when a model version changes. It sits on the balance sheet, it compounds, and it belongs to the company that paid for the work.

Made, not found.