Research & development
We build systems and protocols that help people verify AI claims and reduce harm from overconfidence, manipulation, and unsafe deployment.
Evaluation harnesses
Repeatable tests for reliability, robustness, and generalization. The output is a report, not a vibe.
- Task suites with clear pass/fail criteria
- Failure mode tracking and regression tests
- Uncertainty reporting and calibration checks
Governance scaffolds
We treat governance as an engineering discipline: define gates, log decisions, and document incidents.
- Deliberation-before-action patterns for high-stakes tasks
- Capability gating and staged rollouts
- Audit logs and reproducible artifacts
What success looks like
A successful result is a public artifact: protocol + results + failure cases + a clear explanation of what was learned.
We will publish failures.
If a system fails a safety gate, that’s data. Hiding it is how ecosystems get corrupted.