The dataset nobody recorded

Start with what a decoder actually needs: examples of brain signal paired with the words spoken during it, aligned in time, person after person, session after session. Now go looking for that data in the places everyone points to. Hospital EEG archives are enormous, and they are enormous in the wrong direction. The clinical questions were epilepsy, sleep disorders, stroke. The recordings show seizures and sleep. Language was never the question, so the pairing was never recorded.

The academic results confirm the shape of the hole. Meta's Brain2Qwerty work is the honest benchmark: with MEG, a machine few institutions can house, sentence decoding reaches about 61 percent word accuracy on average, 78 for the best participant. With wearable EEG, character error rates sit near 65 percent. Roughly a third of characters survive. That is what the state of the art does on data that was never designed for the purpose.

Hours of brain recording exist. Hours of brain-plus-speech barely do.

Hours are the unit

Foundation models are bought with hours. MANAS-1 turned sixty thousand hours of EEG into a 400-million-parameter model that transfers across recording setups. On the neural-plus-language side, the largest number we can verify from a competitor is Conduit's reported ten thousand hours. Set against how language models are trained, these are rounding errors, and the paired subset is the smallest of all. Whoever collects paired hours fastest holds the ingredient everyone else has to rediscover.

The languages nobody is collecting

There is a second gap inside the first. The field's data is overwhelmingly English. Indian languages and Arabic are spoken by billions of people and backed by essentially no paired datasets. For a company starting its collection now, that is not a deficiency to apologize for; it is the choice of where to dig. We dig where the shelf is empty, and where the resulting model serves populations the incumbents have skipped.

So Phase 1 is a study, not a product launch. Consent, ethics approval and de-identification come before the first recording. The moat is not a trick; it is the compounding of hours that no archive can lend us and no competitor can buy, in languages no competitor is collecting. That is a slow sentence to write and a fast one to defend in diligence.