Independent Research · Est. 2025
CIRWEL Systems — A research preface, with running code
Researchers study agents at two moments — before deployment, and after an incident. In between, while agents are actually running, the behavior that matters most happens where you can't see it: the slow drift, the approach to failure. Traces are abundant, but nothing reads an agent's own state while it acts. UNITARES makes that window measurable — a continuous, class-calibrated reading of each agent's state, with signed provenance behind every intervention.
One principle — nothing inferred. State is measured, identity is proven, memory is sourced, every action carries a recorded reason. Never reconstructed after the fact.
— The receipts
— Production snapshot
Measured 2026-08-22, refreshed daily against the production fleet, in continuous operation since November 2025. Single-operator self-traffic — our own fleet, not external adoption.
— List of plates
§01 — The thesis
Operating a heterogeneous agent fleet in continuous production surfaces a class of failure that pre-deployment evaluation and post-incident forensics don't catch: the slow, silent drift of agent state across hours and days of normal running. The traces show what the agent did, abundantly. They don't show what state it was in while doing it — because nothing on the agent or alongside it is producing that reading.
Logs are what an agent did.
Self-state is what it was while doing it.
CIRWEL's response is a runtime layer that gives each agent a continuous reading of its own state — a four-dimensional vector summarizing capacity, signal integrity, uncertainty, and the imbalance among them, updated from every check-in. The reading is calibrated against agents of the same class (a coding session, a research conversation, a resident cron, an embedded service, an ephemeral parser), because a long-running coding assistant does not behave like an ephemeral parser, and neither behaves like an embedded service. Drift is then detected against the right reference, not an averaged one.
The framework is described in a paper and has been governing CIRWEL's own development fleet continuously since November 2025.
— A measurement on our own fleet
On a 30-day slice of our own production data — 13,310 governance observations across the fleet — replaying each decision with per-class baselines instead of one fleet-wide baseline produces a different verdict 28.9% of the time. 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 measured on our own production data. The gap is large enough to show that averaging dissimilar agents into one distribution is not a benign default; class-conditional calibration is the response.
The instrument is held to the same bar. Where a result is negative, it is published as negative: none of the seven candidate state streams beat a last-value persistence baseline at predicting negative task outcomes, and the confirmatory re-read is pre-registered for December 2026 with a kill criterion. A monitor that cannot lose is not a measurement.
§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 — caught while it is happening, early enough to intervene, late enough to be evidence-based.
→ § how the layers fit togetherii
One provenance-tracked graph the whole fleet reads and writes — discoveries, decisions, and dialectic resolutions, queryable by the operator and by the agent mid-task. Nothing the fleet learns lives in one head; nothing it acts on is unsourced. Postgres with the Apache AGE graph extension and pgvector retrieval, fused by reciprocal-rank ranking.
→ § the shared substrateiii
Identity is process-bound, not a claimable label: fresh agents mint fresh identity, lineage is declared rather than inherited, and every intervention carries a signed chain back to the observation that triggered it. Which process acted is provable; what it claims about itself is recorded with its source, not taken on faith — replayable by a regulator, an underwriter, or the next-shift human.
→ § identity & lineage§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 embedded edge service, and a Discord bridge — has been governed continuously by UNITARES since November 2025.
Each check-in resolves to a binary action — proceed or pause — qualified by a sub-action (approve or guide under proceed; reject and other reasons under pause), together with a signed lineage back to the observation that triggered it.
Living under one's own framework is the lowest-cost 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.
The open artifacts: a 32,181-trajectory EISV dataset, distilled teacher and student models, and a verdict-counterfactual reproducibility kit at huggingface.co/hikewa. The calibration, drift, and provenance primitives are Apache-2 — meant to be built on, not reimplemented.
