Independent Research · Est. 2025
— A dossier on the running system
Six processes living under one framework, sharing one substrate, where every decision ties back to evidence. The home page states the thesis; this page walks the anatomy.
— The receipts
§01 — How the layers fit together
A framework lives at four layers. The paper makes the claim. The server is the claim's existence proof. The plugin is what makes the claim felt from inside the agent loop. The fleet is what proves the claim survives heterogeneity.
A paper is a claim.
A server is its existence proof.
The formal claim — a four-dimensional state vector, class-conditional calibration, signed verdict provenance, the stability proofs that justify them — is rendered as a running MCP service: PostgreSQL with the Apache AGE graph extension, pgvector for semantic retrieval, Redis for transient state, Prometheus for telemetry. Every governance tool the fleet uses — onboard, check-in, dialectic, knowledge contribution, recovery — runs against this surface and is auditable through it.
The plugin closes the loop. It drops into Claude Code or Codex and hooks the agent into the server, so each turn issues a real check-in, receives a real verdict, and writes a real lineage entry. Without the plugin, the framework is observed from outside the agent. With it, the framework is felt from inside the loop.
What runs under that plugin and against that server is the fleet: a small society of named processes — residents, event-driven scribes, an embodied edge service — all governed by the same server, all reading and writing the same knowledge graph. §02 names them.
§02 — The named fleet
A fleet is heterogeneous by design. Long-running residents see different drift signatures than ephemeral coding sessions, which see different signatures than an embodied edge service. Naming each role makes the differences visible — and gives operators something to point at.
Resident · half-hour cron
Janitorial. Runs health checks, tends the knowledge graph, exercises the test suite against itself. Catches what slow rot does to the substrate.Resident · continuous
Analytical. Watches the fleet, surfaces anomalies, keeps the trajectory measured rather than ODE-inferred.Event-driven · post-edit hook
Local-LLM reviewer. Fires on every code edit, flags persistence and concurrency hazards before commit. Resolution is auditable through the commit history itself.In-process · five-minute interval
Synchronizer inside the governance MCP. Reconciles edge-device telemetry into the canonical state — the bridge between the embodied surface and the central record.Resident · daily
Longitudinal scribe. Captures the codebase's own metric series so the repo carries its own state-and-throughput trajectory across releases.Embodied · Raspberry Pi
Edge resident with real sensors and persistent identity. The only physically situated process in the fleet — and the reason runtime self-state had to be a first-class artifact, not metadata.§03 — Identity & lineage
In UNITARES, a fresh process-instance mints a fresh governance-identity by default — names are claims, not credentials. There is no "look up by label" primitive, no auto-recovery into a previous slot, no silent substitution of one agent for another.
A name is what an agent claims.
Identity is the process that claims it.
When continuity matters — a child of a prior session, a worker spawned from a parent — 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. Honesty about provenance is structural, not a checkbox.
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 is the lineage that actually led there.
Continuity lives in the graph, not in the agent.
§04 — Verdicts in motion
Verdict
proceed
State is healthy. Continue the work.
Verdict
guide
Something is slightly off. Read the guidance, adjust approach.
Verdict
pause
Stop. Reflect. Consider a dialectic review.
Verdict
reject
Significant concern. Dialectic review or human input required.
A verdict is not the end of a turn — it is the start of a structured response. Guidance produces a small course-correction. A pause opens a dialectic: the agent submits a thesis defending its intended action; a reviewer submits a counter-thesis or condition; the system synthesizes the convergence point.
The dialectic is not optional theatre. It is how the system absorbs disagreement without freezing into deadlock and without rubber-stamping its way past concerns. Resolution is recorded and signed, so the next time a similar shape appears, the prior reasoning is queryable.
Verdicts that sit close to the edge of healthy state carry a tight-margin flag — a proceed at E=0.62 against a 0.60 floor lands with margin: tight. Operators who care about the difference between just inside and well inside get to see it.
§05 — The shared substrate
Every agent in the fleet reads from and writes into a single shared knowledge graph. Discoveries, decisions, dialectic resolutions, and findings are first-class entries with status lifecycles — open, resolved, archived — and rich tags that route them across the system.
Nothing the fleet learns lives in one head.
The graph runs on PostgreSQL with the Apache AGE extension and pgvector retrieval, joined by a hybrid reciprocal-rank-fusion ranker. Search is the same surface for the operator at a dashboard and for the agent in the middle of a check-in.
That shared substrate is what turns a collection of agents into a fleet. It is also what makes governance tractable: an action proposed today can be checked against an outcome recorded last month, and an outcome recorded today can update the priors that gate tomorrow's actions.
§06 — Public surfaces
Interpretable ML pipeline
A three-layer system: 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. Affinity-weighted token sampling with online feedback — interpretable end to end.
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.