EyeTrustAIRisk guarantees in a new context
Transfer error control to new deployment environments while distinguishing errors among released outputs from errors across all inputs.
← ResearchStatistical guarantees
Research agenda
Statistical guarantees
for deployed, adaptive AI.
Our program: closed-loop conformal risk control. We study how guarantees can remain meaningful as environments, decisions and feedback change.
Active research directions. These questions are not claims of completed results or currently available product guarantees.
I / Across environments
When does a guarantee travel?
A model may be reused across populations, institutions and tasks. Its evidence needs a defensible scope.
Learning across environments
Use evidence from many settings to understand when a new setting can inherit statistical support.
Validity under drift
Determine when a guarantee needs to expire, when new evidence is required and how recalibration should restore support.
Changing the definition of error
Preserve validity when teams refine error definitions, subgroups or risk limits after seeing data.
II / Within adaptive systems
How do guarantees survive feedback?
Deployed AI makes sequences of decisions, receives incomplete feedback and can change what happens next.
Multi-step workflows
Allocate risk across a sequence of actions so that guarantees concern the whole workflow, not only isolated predictions.
Selective and delayed feedback
Account for labels that arrive late or are observed mainly because a case was selected for human review.
Active auditing
Choose which cases to review to strengthen future risk control without compromising current guarantees.
Systems that change their environment
Study risk when the actions of an AI system alter the population and outcomes it will encounter.
The connecting question
Can guarantees survive a closed loop?
Our aim is a coherent theory connecting changing environments, adaptive policies, selective feedback, certificate expiration and recalibration. Medical imaging, autonomous systems and AI agents provide settings in which to test that theory.
A shared scientific challenge.
Interested in these questions? We would like to hear from you.