Introducing Hokusai

Every coding harness generates the same exhaust: which model got picked for which task, and whether the output worked. That exhaust is the most valuable training signal in the industry right now. The question is who ends up owning it.

Right now the answers all look the same in the end. The labs ship internal routers that capture the choice signal from every API call. The harnesses, as they scale, reach the same conclusion the labs did — that "we know which model is best for which task" is the most valuable thing they sit on top of, and the move is to wall it off. Cursor's $60B tie-up with xAI is the clearest tell so far, but it won't be the last. The structural pressure on every harness that gets big enough is to become a value-capture layer in its own right, then sell that to the megacorp with the largest check.

Hokusai is the alternative for people who don't want it to end that way. The pitch is narrow and concrete. The routing signal — which model, which task, which outcome — should be owned by the engineers and integrators who produced it, not by whichever company acquires the harness they happened to ship inside this year. And the protocol that aggregates that signal should keep running on the same terms tomorrow as today, regardless of which board meeting changes which strategy. The first product is a task router: drop-in middleware that picks the right model for each coding task using a shared learned policy, and mints tokens to the integrators whose outcome data improves that policy. The choice layer sits above gateways — gateways move the call, we pick which call to make. Same factory applies to tool selection, retrieval policy, and code-review critics. Tokens are a position in the router, not a speculative asset. Hold them, redeem them for USDC, or use them to govern how the policy evolves. The design routes financial value back to the integrators whose outcome data improved the policy — that's the whole point of having a token in the loop rather than a closed margin pool. We're not going to tell you what that's worth, because nobody honestly can, but we will tell you that the mechanism exists and that it pays the people who did the work. Over the next few weeks we'll publish the mechanics: how DeltaOne measures marginal improvement, how outcome logs stay auditable without giving up the eval suite, and what the SDK surface looks like for builders integrating today. If you've been watching the harness layer consolidate and wondering whether there's a version of this that doesn't end with one logo owning the signal, talk to us.