RESEARCH DIRECTION
Capability without alignment is a liability. We treat safety as an architectural discipline: agents whose scope of action is explicit and bounded, whose uncertainty is calibrated and surfaced, whose reasoning is inspectable, and whose deployment is gated on demonstrated understanding rather than demonstrated performance.
A system that overstates its confidence, hides its uncertainty, or optimizes appearances over substance is unsafe long before it is powerful. We build anti-deception measures into evaluation itself — proper scoring, adversarial probes, and audits that make confident nonsense costly — and apply the same standard to our own research claims.
We prefer agents whose action space, resource envelope, and escalation paths are architectural facts rather than behavioral hopes. Permissions, audit trails, and fail-safes are designed in from the first commit; oversight is real, and a human can always reconstruct what the system did and why.
A well-calibrated system knows where its competence ends — and the safe response at that boundary is deferral, not improvisation. We study uncertainty-aware decision policies that trade autonomy for oversight exactly when the model's own error bars say they should.
WORKING PRINCIPLES
CONTINUE EXPLORING