Research

The studies behind Unl. The first two were frozen before they ran and published whichever way they fell, each with its protocol beside the result.

The first pass, measured.

When a decision changed partway through the work, agents with Unl needed 3 corrections in 36 tasks. Agents reading a decisions file nobody had updated needed 14, and agents with no file needed 18. Against a file someone kept perfectly, Unl came out even, with nobody keeping it. Frozen before the run and published whichever way it fell.

Read the result · The protocol · Every row

Measured on real fleets.

With Unl, the fleets acted within everything decided on 70 of 90 tasks against 61. Pre‑registered and hashed.

Read the result · The protocol · The same test as Unl ships

How little it had to say.

In a week of real turns in a coding agent, Unl held 343 decisions and handed over a median of 286 tokens a turn. One decision arrived in full and the rest as a single line each, for the agent to open when it needed them. Nothing the agent already had was sent twice.

For what these mean for building with Unl, read the developer page.