The world is not only software.

Models are trained on everything ever written. But most of what happens on Earth is never written down. It passes through instruments as signal, then disappears. Almost none of it has ever taught a model anything.

We turn that signal into training data, so labs can build models that are better at science.

Measurement is a better teacher than text. Written evaluations drift into the training set and get optimised against. A measurement doesn't care what the model expected. It can only be satisfied by learning how the world actually behaves.

A model trained this way should be able to read an instrument, predict what it will show, and eventually operate it.

We're backed by Y Combinator, currently in the F26 batch.

We're not hiring and won't reply to applications or CVs.

wilhelm@qokedas.com San Francisco

The market for measurement

Frontier labs have made training data one of the largest line items in AI. Mercor, the biggest supplier of human data work, did $614M in the first half of 2026 and is running at a $2B annual pace, with roughly 91% of that revenue coming from frontier labs. Anthropic reportedly budgets on the order of $1B for RL environments alone. The money has moved to whatever grades the model.

It moved there because verification, not model capacity, has become the bottleneck. When a grader is wrong, the model learns the wrong thing: 28.5% of SWE-bench Verified tasks accept incorrect patches, and models score about 14 points higher on exactly those tasks. Better graders measurably produce better models, which makes grader quality a first-order training input rather than an evaluation detail. DeepSeek-R1-class reasoning came from little more than a prompt and a verifier.

Measurement is the grader that cannot be wrong in that way, and the precedents for training against it are the strongest results in scientific ML. AlphaFold came out of CASP, a contest that scored blind predictions against withheld experimental structures: one contest, one Nobel. GraphCast, trained on weather measurements, now out-forecasts the physical simulators that generated its training data. Machine-learned force fields reach near quantum-chemistry accuracy at roughly a thousand times the speed.

Yet frontier models still fail most research-level scientific coding and physics benchmarks, most scientific data is never published at all, and much of civilization's validated physics lives in decades-old code maintained by a shrinking community. The infrastructure is starting to catch up: Anthropic's Model Hardware Standard, released in August 2026, is a spec for agents operating laboratory instruments, piloted at Genentech, Janelia, and QuEra. The signal layer of the world is being connected to models. Someone has to supply the data.