Independent Research · Est. 2025
CIRWEL Systems — A dossier on the running system
Four answers, one substrate: who an agent is, what the fleet knows, what state each agent is in, and why every action was allowed. The home page states the thesis; this page is the architecture.
— The receipts
§01 — How the layers fit together
A claim, its existence proof, the thing that makes the claim felt from inside the agent loop, and the proof it survives heterogeneity.
The formal claim — a four-dimensional state vector, class-conditional calibration, signed verdict provenance — runs as a live MCP service: PostgreSQL with the Apache AGE graph extension, pgvector retrieval, Redis for transient state, Prometheus for telemetry. Every governance tool runs against this surface and is auditable through it.
The plugin drops into Claude Code or Codex and hooks the agent into the server, so each turn issues a real check-in and receives a real verdict. The four sections below are the substrate that makes that possible: who acted (§02), what the fleet knows (§03), why an action was allowed (§04), and the heterogeneous workloads it all governs (§05).
§02 — Identity & lineage
A new process mints fresh governance-identity by default. Names are claims, not credentials.
There is no "look up by label," no auto-recovery into a previous slot, no silent substitution of one agent for another. When continuity matters — a child of a prior session, a spawned worker — lineage is declared via a parent reference, not inherited by claim. The identity API exposes the source of every label, so an auditor can see whether a name was operator-assigned, self-claimed, or system-generated.
The cost is friction at session start. The benefit is that the chain of custody behind every verdict is real: the agent that produced an outcome is the agent the system thinks it was, and the lineage that led there actually led there.
§03 — The knowledge graph
Every agent reads and writes one provenance-tracked graph. Nothing it learns lives in one head; nothing it acts on is unsourced.
Discoveries, decisions, dialectic resolutions, and findings are first-class entries with status lifecycles — open, resolved, archived — and tags that route them across the system. It runs on PostgreSQL with the Apache AGE extension and pgvector, joined by a hybrid reciprocal-rank-fusion ranker. Search is the same surface for the operator at a dashboard and the agent mid-check-in.
That shared substrate is what turns a collection of agents into a fleet, and what makes governance tractable: an action proposed today is checked against an outcome recorded last month, and an outcome recorded today updates the priors that gate tomorrow's actions.
§04 — Verdicts in motion
Action
proceed
Action
pause
The decision is binary — proceed or pause — qualified by a sub_action. Separately, each cycle carries a thermodynamic health reading (safe / caution / high-risk) derived from the state vector, not the decision. A pause opens a dialectic: the agent defends its intended action, a reviewer submits a counter-thesis or condition, the system synthesizes the convergence point — recorded and signed, so the next similar shape is queryable.
Alongside the decision, each cycle reports one proprioceptive margin — settling / comfortable / tight / warning / critical — measuring headroom to the nearest of three policy boundaries (risk ceiling, coherence floor, void ceiling), with a companion nearest_edge. A proceed at risk 0.58 against a 0.70 pause threshold lands margin: tight, nearest_edge: risk — not because one dimension crossed a floor, but because distance to the nearest boundary fell inside the tight band.
§05 — Governed workloads
Class-conditional calibration only works if the classes are real. CIRWEL's own deployment spans six kinds of process, each with a different state signature — the heterogeneity the calibration is tested against. The monospace tags are the running process identifiers.
30-min cron · vigil
Health checks, knowledge-graph tending, and a self-test against the suite. Catches what slow rot does to the substrate.Always-on · sentinel
Watches the fleet and surfaces anomalies, keeping the trajectory measured rather than inferred.Event-driven · watcher
A local-LLM reviewer that fires on every code edit and flags persistence and concurrency hazards before commit.5-min interval · steward
Runs in-process inside the governance service, reconciling edge-device telemetry into the canonical state.Daily · chronicler
Captures the codebase's own metric series so the repo carries its state-and-throughput trajectory across releases.Raspberry Pi · real sensors
The only physically situated process — real sensors, a stable cross-session identity, and the reason runtime self-state had to be a first-class artifact rather than metadata.§06 — Public surfaces
Interpretable ML pipeline
Rule-based generator → LoRA teacher on Qwen3-4B → distilled RandomForest student for Pi deployment. The trajectory dataset (32,181 records, nine shape classes) and the teacher and student models are public on huggingface.co/hikewa.
Operator surface
A standalone bot that surfaces governance events, agent presence, and autonomous actions as a living server — the operator's window into a fleet that does not stop when the human does.