Independent Research · Est. 2025
— A research preface, with running code
Most AI safety reasons about agents in two windows: pre-deployment evaluation and post-incident forensics. Between them — while agents are actually running — there is a measurement gap. Continuous, class-calibrated self-state telemetry, with signed provenance behind every intervention, fills it.
— The receipts
§01 — The thesis
The discipline of AI safety has built two strong instruments and one large gap. Pre-deployment evaluation tells you whether a model behaves on a benchmark. Post-incident forensics tells you what went wrong after it didn't. In between — across the hours, days, and weeks an agent is actually running — we have largely been guessing.
Logs are what an agent did.
Self-state is what it was while doing it.
CIRWEL Systems builds the runtime layer that closes the gap. Each agent carries a continuous, four-dimensional self-state vector. The vector is calibrated against a baseline specific to the agent's class, because a long-running coding assistant does not behave like an ephemeral parser, and neither behaves like an embodied service. Drift is detected against the right reference, not an averaged one. Every governance verdict — proceed, guide, pause, reject — carries a signed lineage back to the observation that produced it.
The framework is described in a paper, covered by nine provisional patents, and has been governing CIRWEL's own development fleet without interruption since November 2025.
§02 — Three pillars
i
A coding agent and a research agent are not held to the same statistics. UNITARES learns separate baselines per agent class from production telemetry, so drift in one class is not masked by noise from another.
→ § how the layers fit togetherii
Continuous state observation catches behavioral drift while it is happening, not in the post-incident review. Verdicts arrive early enough to intervene, late enough to be evidence-based.
→ § verdicts in motioniii
Every intervention carries a signed lineage back to the observation that triggered it. Regulators, underwriters, and the next-shift human can replay the chain — not just read a verdict.
→ § identity & lineage— A first measurement
On a 30-day slice of production data — 13,310 governance observations across the fleet — replaying each decision with per-class baselines instead of one fleet-wide baseline disagrees on 28.9% of them. The disagreement skews systematically: state vectors the fleet-wide baseline classifies as healthy or borderline are usually flagged as drifting under per-class baselines. Per-class flip rates range 15–33%.
This is the cost of "average all agents into one distribution" making its first quantitative appearance, on real production data. Class-conditional calibration is the response. (§11.6 of the paper.)
§03 — In production
The system you read about on this page also wrote, tested, and shipped a meaningful fraction of itself. CIRWEL's development fleet — a heterogeneous mix of long-running resident agents, short-lived coding sessions, an embodied edge service, and a Discord bridge — has been governed continuously by UNITARES since November 2025.
Living under one's own framework is the cheapest credibility a research operator can offer. We treat it as the floor, not the ceiling.
This page is part of the loop. The colophon below shows the exact commit and build time that produced what you are reading.
§04 — Engage
The server at CIRWEL/unitares boots locally. The plugin at CIRWEL/unitares-governance-plugin drops into Claude Code or Codex and emits real verdicts against it.