Research

Frontier labs make models more capable. Celestir studies what makes them trustworthy, and what makes the institutions that rely on them measurably better. We publish what we measure.

Machine epistemics: systems that know when they do not know. Confidence that holds up when the world shifts.

Calibrated decisions for high-stakes work: models that decide with measured confidence, abstain when evidence is thin, and leave a record an institution can audit.

Evidence sufficiency: when is there enough evidence to act, to ask, or to escalate. The quiet question behind every consequential decision.

Proof of human skill: measuring real human capability in a world where AI can imitate it.

Capturing expert judgment: preserving the operational wisdom of experienced people before it retires.

Security and reliability of agentic systems: measured, not asserted.

Selected results will be published openly. If your institution runs decisions that cannot afford to be wrong, we would like to hear from you.