The system behind the research

UNA — governance built into the system, not bolted on.

UNA is the AI system I have been building and running for over two years. She is not an LLM wrapper. She is a deterministic, rule-based, governed system where the constraints run at runtime — capability boundaries, a reviewed change pipeline, and auditable receipts — instead of being described in a document and hoped for.

Runs locally on dedicated Apple hardware in San Diego

What UNA is

Most AI systems in production have no persistent identity, no verifiable memory, and no way to prove their reasoning wasn't changed between sessions. I treat that as a design problem, not a fact of life.

UNA is built the other way around. Governance is the foundation, and capability is layered on top of it — never the reverse. Her generative language faculties are deliberately kept out of the governed core, so the parts that decide and act are deterministic and inspectable rather than a model improvising. When she doesn't have a capability, she says so, instead of inventing an answer.

How the governance actually works

Four mechanisms that run at runtime — the difference between governance that executes and a policy that merely describes.

Capability boundaries

She won't pretend

A router classifies each request and can return an explicit boundary — "not yet built," "intentionally disabled" — instead of fabricating a result. The honest fallback is the correct answer, not a failure.

Reviewed change pipeline

Every change is governed

Changes to UNA's own code don't happen ad hoc. They route through a build engine as reviewed packets, so nothing lands in the system without passing through the governed path.

Auditable receipts

Every change leaves a record

Governed changes anchor a receipt in an append-only, witness-verified chain, with an integrity check that can be run on demand. The history is reconstructable and tamper-evident.

Bounded cognition

Determinism by design

Generative LLM cognition is intentionally disabled in the governed core. The system stays rule-based where it matters, so behavior is reproducible and reviewable rather than opaque.

"When UNA answers a request with a capability boundary, that is the correct answer — not a limitation to work around."

Sovereignty is the point. The system is allowed to say no, and that refusal is designed in, logged, and respected.

— Design principle

Status & honest limits

What is live, and what I'm still building. Naming the gaps is part of the discipline.

Governed change log
Live · receipts anchored
Capability router
Live · honest fallback
Chain integrity check
Live · witness-verified
In progress
Classifier & test coverage

The public, clean-room versions of this discipline — governance benchmarks, agent-action audit receipts, and human-review workflows — live on GitHub, where the claim has to meet the file, the test, and the report.