Unified demonstration · three repositories, one flow · self-contained build
Governance you can run, not just describe.
Most AI governance is a document. This is the same policy expressed as code you can execute: ask a question, watch it get grounded or refused, then watch a deterministic gate issue a verdict and route it to a human. Every step leaves evidence.
The machinery is the easy part. Governance is the set of questions the machinery must answer — what the system is allowed to do, who decided, where a human must stand, and how you prove it to someone who wasn't in the room. Every page here answers one of them, and the subject of every answer is a person.
This edition is a single HTML file — no backend, no build step, no external services, no model in the loop. Every behavior on every page is deterministic code running in your browser. Open it from a file share or host it as one static page.
The 90-second version: ask → verdict → your sign-off → the audit trail. You play the human in the loop. The production day streams ~200 decisions through the same gate — automation does the routine; you work the exceptions.
The governance story — five questions, five answers
Strip away the machinery and governance is five questions with named owners. Each page of this demo exists to answer one of them.
What this is assembled from
Two working repositories and one framing repository, joined into a single control surface.
- governed-playbook-assistant — governance on the way in: retrieval, citation binding and refusal thresholds.
- ai-decision-governance-gate — governance on the way out: seven deterministic controls producing PASS / REVIEW / BLOCK.
- framework-optimization-lab — the framing: controls that are cheap to run and hard to argue with.
Governed assistant
Retrieval is deterministic keyword search over a synthetic operating playbook. If nothing is relevant enough, the answer is refused before any drafting happens. Turn governance off to see the same question answered with confident, unsourced figures.
In this self-contained build the drafting layer is extractive code, not a model — answers are assembled from the retrieved passages themselves. A production deployment swaps in a language model for drafting; every control around it is identical.
Governance question 1 — what is this system allowed to do?
The boundary was set by a person, not learned by a model: the assistant may answer only from the approved playbook, and below the relevance threshold it must refuse. The refusal happens before any drafting — an unanswerable question never becomes a confident answer.
Policy-as-code gate
The drafting layer proposes; the gate decides. Load a scenario, change any field, and watch the verdict move. Nothing here is probabilistic — the same record always returns the same verdict.
Governance question 3 — what risk does each control manage?
Every control exists because of a way this could hurt someone: invented sources, unsourced figures, false precision, a high-risk decision nobody owns. The gate doesn't enforce style — it prevents named failures, and the same record always gets the same verdict, which is what makes it defensible.
Verdict
Run the gate to see all seven controls and the resulting verdict.
Review queue
REVIEW is not a warning to be dismissed — it is a stop with a named owner. BLOCK never appears here: a blocked output must be fixed at the source, not signed away.
Governance question 4 — where must a human stand, and why there?
Here — because this is where consequences land on people. A machine may propose, but an uncertain output waits for a named person who signs with their name, their role, and their reasons. And some things no signature can release: a BLOCK is fixed at the source, not approved away.
Reviewer identity
Governance dashboard
Everything below is generated by your own session — the same evidence an assurance function would ask for. Nothing is pre-baked. Records live in this browser only.
Governance question 5 — how do you prove it to someone who wasn't in the room?
With a record you didn't have to reconstruct. Every decision below was written down as it happened — attributable, append-only, downloadable. This is the difference between "we have standards" and "we can show you."
Decision trail
Audit trail
Append-only; newest first. Every step is attributable.
Playbook and evidence registry
Governance depends on a bounded corpus. The assistant may answer only from these passages, and the gate rejects any citation that is not in the registry below. All content is synthetic.
Governance question 1 — who decided what the system may know?
A person chose this corpus and registered this evidence — accountability starts with a boundary someone owns. Notice the playbook's own rule: every decision it governs must name exactly one accountable human. The system inherits its governance from the people who bound it.
Operating playbook
Evidence registry
A citation outside this list is treated as fabrication.
How it works
The split matters more than the stack: a drafting layer proposes prose, and deterministic code decides. Every verdict in this demo is reproducible without a model in the loop — and in this self-contained build, there is literally no model in the loop.
Governance question 2 — who decided, and by what authority?
Never the drafting layer. Authority lives in code a person can read, thresholds a person set, and tests a person can run — which means every verdict can be challenged, reproduced, and owned. That is what separates a governed system from a well-behaved one.
What the drafting layer does
- Composes prose from retrieved passages only.
- In a production deployment this is a language model under instruction; here it is extractive code.
- It never assigns a verdict, tier, citation or risk level — in either case.
What deterministic code does
- Ranks passages with BM25 and computes the relevance score.
- Refuses below the threshold before any drafting happens.
- Assigns citations, tiers, and the seven-control verdict.
Control map
Each control traces to a function and a test in the source repositories.
Honest limits
- The playbook, evidence registry and figures are synthetic demonstration data.
- Retrieval is keyword-based, not semantic — chosen so refusals are explainable by a number.
- This build has no model: answers are assembled extractively from the retrieved passages, and the "ungoverned" baseline is a labeled simulation of what an unconstrained assistant does.
- Records are stored in this browser only; there is no identity or access model here.
- Seven controls are a demonstrable core, not a complete assurance framework.