I study how AI agent work stays attributable, reviewable, and verifiable across sessions, runtimes, and principals.
Open to research collaboration and fellowships: founder@cirwel.org.
For UNITARES and how to run it:
cirwel.org ↗.
Papers & preprints
Trajectory Identity
Formulates agent identity through behavioral trajectory statistics—attractor basins, recovery dynamics, and homeostatic regulation—rather than a stored label or static memory, and evaluates the framework across an embodied deployment and a four-agent multi-harness pilot.
Accountability Without a Trusted Center
Designs a testbed and pre-registered evaluation protocol for cross-principal agent governance mediated by verifiable attestation rather than a central authority, and demonstrates offline Ed25519 attestation verification across two hosts.
Reading guide: each paper’s question, evidence, and scope ↗
Runtime signals, their interpretation, and policy effects have distinct evaluation questions. Full research record ↗
Software
UNITARES
Self-hosted infrastructure for identity, claims and evidence, shared memory, longitudinal state, review, policy, recovery, leases, and coordination. Exposes MCP, REST, an SDK, and operator surfaces.
Repository · Paper · PyPI · Archived release
UNITARES Resident
Conversation loops, provider adapters, tools, scheduling, and queues built on the public UNITARES contract. The resident is optional; a custom runtime can use the SDK directly.
Host adapters & plugin
Thin integrations that connect Claude Code, Codex, Hermes, and other hosts to UNITARES check-ins, review, memory, and policy surfaces without moving the agent loop into Core.
Anima
A Raspberry Pi sensor system used to test longitudinal telemetry, continuity, and state-estimation machinery against a physically situated process. Its traces support the digital-proprioception work and the published datasets.
Datasets & reproduction
EISV Trajectories
32,181 labeled trajectory windows: 20,655 overlapping real windows from the embedded deployment, plus 11,526 synthetic examples across nine shape classes.
Verdict Counterfactual Kit
Data and scripts for replaying how governance decisions change under alternative baseline and input assumptions.
Zenodo · Hugging Face · Source
CIRWEL is an independent research programme; operational measurements come from the maintainer's long-running single-operator deployment.