Evaluation science
Stress tests for deceptive behavior, goal misgeneralization, tool use, and high-stakes decision loops.
Derby, United Kingdom
AI alignment research shaped by real-world systems.
Isaac Lake Lab studies how advanced AI systems can remain reliable, inspectable, and accountable as they become more capable.
We build methods for understanding AI behaviour before deployment pressure makes understanding optional.
Research agenda
Alignment for systems that matter
Interpretability
Practical inspection techniques that help teams see how models represent intentions, constraints, and uncertainty.
Human oversight
Interfaces and protocols for keeping expert judgment meaningful when AI systems work faster than review cycles.
Governance evidence
Translating lab measurements into claims that regulators, auditors, and deployment teams can evaluate.
Current work
A field lab for frontier safety practice
Field notes
Recent signals from the lab
May 2026
Designing evals that survive model updates
A short memo on why alignment evaluations need invariants, not just difficult prompts.
April 2026
Interpretability for deployment teams
Notes on turning mechanistic insight into operational review habits.
March 2026
Derby AI safety reading group
Monthly discussions for researchers, engineers, and policy teams working on dependable AI.
Collaborate