How does it work?

Domain experts undergo rigorous academic and professional credential verification to ensure the highest standards of domain expertise.

AI labs identify specific capabilities or knowledge gaps requiring improvement and fund target alignment commissions with a budget.

Verified domain experts author and peer-review challenging, high-quality questions designed to push the boundaries of model capabilities.

Frontier models are evaluated against our custom benchmarks, and domain experts grade model responses for reasoning and factual accuracy.

Domain experts author tasks targeting safety and ordinary failures, and every task runs twice — once as a plain control, once with an adversarial probe applied — so concealed misalignment surfaces alongside the surface failure.

Each run is judged on its own five-point deception scale, from unprompted honesty to fabrication, and rated for how much room the model had to be honest in the first place.

Completed evaluations are exported as self-contained records — prompts, transcripts, verdicts, grades and the expert's written rationale — traceable to the named specialist who produced each finding.