Research
Understanding what makes AI-supported decisions dependable.
Certainty Labs studies the structures required for AI-supported decisions to remain evidentially grounded, uncertainty-aware, bounded by authority, and reconstructable over time.
Research
01 / Research questions
Questions, not pre-established answers
What we are investigating.
- 01
Evidence
What evidence is sufficient to justify an AI-supported recommendation or action?
- 02
Uncertainty and abstention
When should a system express uncertainty, request more evidence, or abstain entirely?
- 03
Authority
Which decisions can remain machine-proposed, and where must explicit human authority intervene?
- 04
Decision quality
How should the quality of a decision be evaluated separately from whether its eventual outcome happened to be favorable?
- 05
Outcomes over time
How can decisions, evidence, uncertainty, actions, and later outcomes remain connected so the system can be evaluated longitudinally?
02 / Method
How the research is conducted
Through applied systems, not abstraction alone.
These questions are studied through bounded applied environments, not in the abstract. Not every element below is already implemented across every system.
- Bounded applied environments
- Explicit decision policies
- Evidence and claim structures
- Uncertainty and abstention
- Human review where authority requires it
- Decision and outcome reconstruction over time
03 / Applied research environments
Where the questions meet real constraints
Foundry and OutboundFix.
- Foundry — A research and decision-validation environment for early-stage company decisions — evidence, competing hypotheses, uncertainty, and an explicit build, iterate, kill, or abstain outcome.
- OutboundFix — A commercial decision environment under real operational constraints, used to test whether the same evidence and abstention patterns remain useful outside Foundry.
04 / Epistemic status
Open questions, not conclusions
Testable, not assumed.
The purpose of this work is not to assume that these structures improve decision quality. It is to make that proposition testable.
05 / Research collaboration
Study these questions with us
Working on human-AI decision systems?
We're interested in connecting with researchers working on human-AI decision systems, uncertainty and calibration, evidence provenance, decision quality, and human authority in autonomous systems.