We work with health systems, standards bodies and regulated enterprises on one question: how do you know an autonomous system did what it was supposed to? Every answer ships as a paper, a repository and a dataset that anyone can run.
Each domain starts from a question a partner has to answer in production, and ends with results that reproduce from a public repository.
QUOKKAGUARD is a single Rust gateway, qfire, that sits in front of any OpenAI-compatible model and decides allow, block or escalate for every request and response. Each module is a standalone paper, a layer in the gateway and an adversarial dataset. Together they implement the controls of HAARF, the regulatory framework written with the Task Force for AI Agents in Healthcare.
Partners define the question with their clinicians, security engineers and compliance teams. We build the testbed, red-team it, and publish the result together, in the open.
A concrete failure mode in your deployment becomes a research question with a measurable outcome.
Open-source harness on local models and synthetic data. No PHI or proprietary data leaves your network.
An adversarial dataset with a full data card, so the defense is measured against attacks, not assumptions.
Preprint, repository and dataset go public on the same day, with your team as co-authors.
Code is Apache 2.0 and datasets are on Hugging Face. Partners can audit every number, and so can their regulators.
Findings feed HAARF and the Proof-of-Control Standard through open working groups, so a result becomes a control others can adopt.
Every experiment reruns on local 7B to 120B models with no paid API keys, so a partner can verify it before relying on it.
Thirteen walkthroughs of the research, one paper at a time: the question, the experiment, and the number that settles it. A faster first pass than the PDF. Each module card under Publications has its own ▶ Watch button.
Quome Cloud on YouTube · “Research papers” playlist
Clinicians, security engineers and regulators defining how autonomous agents are verified before they touch patient care. Home of HAARF.
Visit →Steward of the Proof-of-Control Initiative: open standards for verifying what AI agents actually did, in a form anyone can check.
Visit →