The research program behind UNITARES
Research.
I study what can be read from an agent while it works, and how accountability can hold between principals who share no root of trust. The papers carry the results. The evidence ledger in the server repository carries their current status, and it changes with the data.
The receipts
- Program
- Longitudinal runtime-state measurement for AI agents · begun September 2025
- Researcher
- Kenny Wang · ORCID 0009-0006-7544-2374
- Data
- datasets & models
- Current status
- EVIDENCE_AND_LIMITS.md · versioned with the code
§ IThe papers
- UNITARES: Information-Theoretic Governance of Heterogeneous Agent Fleets
- The system itself: how UNITARES tracks the working state of very different kinds of agents, judging each against its own kind rather than one standard for the whole fleet.
- Trajectory Identity: A Mathematical Framework for Enactive AI Self-Hood
- Proposes that an agent's identity is the pattern its behavior keeps over time, where it tends to settle and how it recovers, rather than an ID or a store of memories.
- A Flat Coherence Score: What a Production-Data Replay Does and Does Not Show About Agent Self-State Gating
- Examines whether the deployed coherence score could actually decide anything at its own thresholds, and what replacing it with a score measured against each class of agent does and does not show.
- Accountability Without a Trusted Center
- Designs an open testbed for holding agents accountable across organizations that do not trust each other, pre-registers how it will be evaluated, and runs its attestation mechanism between two machines.
§ IIWorking together
The code and data are open.
If you work on evaluations, multi-agent systems, or runtime oversight and want to run it on your own agents, I would like to hear from you. Get in touch.