Keeping AI safety research trustworthy and scalable.

Avoiding catastrophe from AI depends on safety research that we can trust and that keeps pace with capabilities. Within a few years, AI agents will likely write most of that research, faster than peer review, funding, and oversight can absorb. Most safety work on sabotage and research quality happens inside one lab’s pipeline. Few treat the scientific literature itself as the surface to secure: a shared record that any model, trusted or not, can write into, and that researchers build on without coordinating.

Averroes Labs is a metascience org that builds infrastructure for that surface. We start with audits for sabotage and legibility. Coordination tools, so researchers can move faster together, come next.

Pilot projects

Each pilot turns existing research into a working tool: a filter that publication platforms could run on AI-written papers.

I. Sabotage audit

Can an audit catch a model that quietly sabotages its own research? We extend Redwood Research’s Auditing Sabotage Bench from code to full submissions (paper plus code) and measure the catch rate.

II. Legibility quiz

Can people still follow AI-written papers? We adapt KnowledgeGain, a measure of how much readers learn from science news, to research papers: a paper passes only if a weaker model, standing in for a human reader, can answer an independent quiz after reading it.

Team: Jason Fantl, founder. AI safety research fellow at the Cambridge Boston Alignment Initiative (2026). Previously built distributed ML systems at Raytheon BBN for three years.
Seeking: Founding research engineers (remote is fine), collaborators, advisors, feedback, and funding. We are not funded yet.
Background: Autonomous researchers will strain peer review, funding, and oversight (LessWrong, 2026)
Contact: contact@averroeslabs.org