What we do
Three capabilities, one habit: check every claim.
Whether a claim comes from a model, an agent or another party, we test it before anything builds on it.
AI agents with safeguards
Agents are useful once they can act, and risky for the same reason. We design agent systems around four bright lines that no agent crosses without a person: spending money, using credentials, sending messages, and publishing.
These limits live in code, not in a prompt. Every claim an agent makes is checked before anything builds on it.
Ask about AI agents →Model validation and machine-learning research
Most models that look great on paper fail in practice, because of data leaks, unrealistic assumptions or luck. We test for each of those on purpose: data split into periods that stay locked until the end, realistic costs and delays, stress tests, and comparison against random baselines.
Most ideas fail these tests. The ones that pass can be trusted.
Ask about model validation →Decentralised coordination research
When agents from different owners compete for the same task, someone has to decide who won. We research neutral ways to order those claims using fast-finality blockchains.
Ask about coordination →Recognise one of these questions?
Tell us what you are working on. A few lines is enough.