Decision layers on Hokusai
Each layer is a shared model that picks something — which model to call, which tool to use, which context to retrieve. Integrators pay a per-decision fee. Contributors hold a position in that fee stream.
The router
Technical Task Router · HTASK
Picks the right model, prompt, and harness for each coding task, based on a shared learned policy trained on real outcomes. The first decision layer built on Hokusai.
| Status | Live |
| Current benchmark | 0.4211 cost-adjusted task success |
| DeltaOne reward | 250,000 HTASK per percentage-point lift |
| Token | HTASK — position in the router's fee stream |
Decision layers we want built
Each of these is a routing, selection, or retrieval problem currently solved with hardcoded rules or absorbed into lab margin. If you've already built one internally — or have a thesis on how to build one — Hokusai's primitives turn it into a shared model that integrators pay to use and contributors own a piece of.
Tool selection
Coding agents call dozens of tools and MCPs. Which one fits a given task is mostly guesswork or hardcoded rules.
A learned policy could route to the right tool based on outcome data from many harnesses.
Propose this layerRetrieval policy
Every agent decides what to pull from a user's context, memory store, or codebase. That decision shapes every downstream call.
A shared retrieval policy, trained on what actually got used in successful runs, beats per-harness heuristics.
Propose this layerCode-review critic selection
Different reviewer models catch different classes of bug. No one's measuring which catches what.
A router that picks the right critic for the diff — trained on real PR outcomes, not synthetic benchmarks.
Propose this layerPrompt policy
Prompt strategies are tribal knowledge. Engineers swap notes; nobody measures.
A shared model for which prompt strategy fits which task, owned by the engineers who improve it.
Propose this layerEarlier reference models
These predate the decision-layer focus and demonstrate the protocol's measurement and reward primitives — DeltaOne, bonding curves, on-chain attribution — in non-routing domains. Useful as proof the mechanics work; not the direction Hokusai is going.