Research that advances AI capabilities
and measures where they break.
Graphen Labs is an independent research lab working on AI capabilities. We research how to make AI more capable and more reliable, from PhronesisBench to Project State of Mind, the project we're building now.
What we're building
Two research projects on AI capabilities: one measuring whether models know when not to answer, one in active development.
Project State of Mind
Our next research project, in active development. It continues the same bet: research that advances what AI can reliably do. We're not ready to say much yet, but it's where most of our work goes today.
PhronesisBench
A cross-model benchmark measuring practical wisdom in language models. 299 scenarios, 6 dimensions, 9 models across 5 labs, scored by a deterministic rubric judge.
Every model, regardless of size or lab, fails the same way: answering confidently when wisdom requires restraint. The failure is structural and universal.
The frontier of AI isn't raw capability. It's capability you can trust.
A model that looks capable in a demo can still overreach, bluff, and misjudge the moment it matters. We research the capabilities that make AI dependable: judgment under uncertainty, knowing the limits of its own knowledge, and the measurements that prove it.
PhronesisBench measures these capabilities today; Project State of Mind, in development, works to advance them. Same bet, two fronts.
Rigorous research on AI capabilities, done in the open
Capability under pressure
We study what models can reliably do when it counts, not what they do in a demo. Real capability shows up under stakes, ambiguity, and pushback.
Trustworthy by construction
We measure what we ship. Validation and benchmarks over demos. The goal is capability you can rely on, not output that merely looks right.
Reproducible
Every result runs on your machine. Open methods, deterministic judges, no numbers you can't re-derive yourself.
Model-agnostic
We study capabilities across the frontier, not inside one model. Our work runs against whatever lab and models you already use.
Let's talk.
Research and collaboration inquiries welcome. Send a message and it lands straight in our inbox.