One HTML file. No backend, no build step, no model, no external dependencies.
The Governance Control Tower is a runnable AI governance demo: a governed
pipeline — grounded retrieval, mandated refusal, a seven-control policy-as-code
gate, human-in-the-loop review, and an append-only audit trail — expressed as
deterministic JavaScript in a single page. Open index.html in any browser, or
host it as one static file.
Run it: open index.html locally, or use the hosted copy on GitHub Pages:
https://framework-optimization-lab.github.io/governance-control-tower/
| Page | What it demonstrates |
|---|---|
| Ask | BM25 retrieval over a fictional operating playbook; answers bounded by retrieved passages with app-generated citations and a confidence tier; out-of-scope questions refused before any drafting, with a numeric reason. An "ungoverned baseline" toggle shows the failure mode being prevented. |
| Gate | Decision records evaluated by seven deterministic controls → PASS / REVIEW / BLOCK with reasons. Four bundled scenarios (one clean, one review-routed, one triple failure, one should-have-refused); every field editable so you can watch the verdict move. |
| Review | REVIEW-routed records queue for a named sign-off with a written rationale. A BLOCK cannot be overridden — it must be fixed at the source. |
| Dashboard | Verdict mix, refusal rate, sign-off counts, and the full append-only audit trail (downloadable as JSON). |
| Playbook | The bounded corpus: 11 fictional playbook passages and the 6-entry evidence registry every citation is checked against. |
Everything governance-critical — retrieval, thresholds, confidence tiers, refusal, the seven controls, review rules, and audit events — is fully implemented, deterministic logic. Because the page runs with no backend and no model, three components are deliberately simple:
| Component | How this file implements it |
|---|---|
| Drafting layer | Extractive — answers are sentences drawn verbatim from the retrieved passages. No language model is involved. |
| Storage | Browser localStorage. Nothing leaves the page. |
| Ungoverned baseline | A labeled deterministic simulation of an unconstrained answer, shown to illustrate the failure mode the gate prevents. |
The seven controls implement the governance model enforced by
ai-decision-governance-gate;
the retrieval, confidence tiers, and refusal behavior follow the governance of
governed-playbook-assistant.
All data is fictional and domain-neutral. Clean-room demonstration only.