Across the AI stack.
We study AI where it meets real work: how models find and fix vulnerabilities, how agents use tools, and when a small model is enough. Each area lists the projects behind it.
Agent systems
The context, tools and checks around a model that takes action.
Applied AI
Models that turn documents and questions into results you can check.
AI infrastructure
The code search and web retrieval tools that agent work runs on.
AI security
How models find and fix vulnerabilities, how agent systems are attacked and defended, and what a network fingerprint gives away.
Why we measure first
A claim about a model is a prediction: on this kind of input, it will do this. The only way to test a prediction is to run it on fixed items, with the setup written down, and count.
Models are sensitive to small changes, such as the order of the options or the sampling temperature. A number without its setup cannot be reproduced, and a demo shows one run, not a rate. So we fix the protocol before the result, report a plain baseline next to every model, and link each number to the run behind it.
It pays off. On our first Decision Bench run, a simple model trained on 10,003 labelled questions beat both zero-shot decision models on a 77-label intent task. We would not have known without measuring.
Teardowns, reports and measured runs.
Teardowns of agent systems, a source-by-source review of open-weight cyber capability, and a track on small models that make calibrated, typed decisions.
RESEARCH GRAPHSmall decision models, as a growing graph.
Each paper ends in open questions, and those questions become the next papers.
To see the ideas in working code, start with the open source projects. The lab journal covers what we are building and learning as we go.
Read the lab journalPublished research07
Earlier notes, preserved with their original publication dates.