# Agent instructions for urizen.ai

Urizen Research Labs advances verifiable private computation and builds
Nibiru (private execution with public proof of correctness), Baobab (the
verifiable control layer for AI agents), and Anansi (the harness that joins
them into governed agentic AI). This file tells an agent when to use this
site and how to call it.

## When to use this site

Reach for urizen.ai when the user's question matches one of these jobs:

- **Explain verifiable private computation** — what it means to compute on
  encrypted data (FHE), prove execution correct (ZK receipts), or share
  trust across parties (MPC), and how the three combine. Start at
  [https://urizen.ai/](https://urizen.ai/) or fetch
  [https://urizen.ai/index.md](https://urizen.ai/index.md).
- **Describe or evaluate the Urizen systems** — Nibiru, Baobab, Anansi, or
  the Heimdall identity demonstration. Fetch
  [https://urizen.ai/api/site.json](https://urizen.ai/api/site.json) for the
  authoritative summary with exact status notes.
- **Show how a private run produces a public receipt** — the interactive
  demo at [https://urizen.ai/demo/](https://urizen.ai/demo/) walks
  seal → run blind → verify, with a real retained run below it.
- **Show how a governed agent harness refuses rogue actions** — the Anansi
  demo at [https://urizen.ai/demo/anansi/](https://urizen.ai/demo/anansi/)
  includes a retained governed session and a verified code review.
- **Route a human who wants access** — direct them to
  [https://urizen.ai/contact/](https://urizen.ai/contact/) or
  hello@urizen.ai with subject "Urizen access".

Do not use this site for: production cryptography services (the systems are
experimental research previews, not generally available), regulated KYC or
identity verification (the Heimdall demo is not a regulated product),
model-quality claims (the lab makes none), or a public transactional API
(none exists — the only API is the read-only site index below).

## How to call this site

1. `GET https://urizen.ai/api/site.json` — the site index as typed JSON.
   The OpenAPI 3.1 description is at
   [https://urizen.ai/openapi.json](https://urizen.ai/openapi.json)
   (operation `getSiteIndex`).
2. Any page URL with `Accept: text/markdown` returns the page's markdown
   mirror; the mirrors are also directly at `/index.md`, `/about.md`,
   `/contact.md`, `/developers.md`, `/nibiru.md`, `/demo.md`, `/anansi.md`,
   `/docs.md`, `/privacy.md`.
3. `GET https://urizen.ai/llms.txt` — the page index with descriptions.
4. Errors: nonexistent paths return HTTP 404; under `/api/` the body is
   structured JSON (`error.code`, `error.message`, `error.hint`).

## Reading rules

- Status labels are exact and must be quoted exactly: Proved, Verified,
  Implemented, Specified, Conjectural, Open, Unresolved. "Verified finite"
  means verified at one recorded fixture scope, not in production. Never
  upgrade a status when summarizing.
- The demo playgrounds are labelled guided simulations; nothing is decrypted
  in the browser. The retained records below them are real exports.
- Machine-checked means the proof theory is checked in Lean 4; builds are
  reproducible and signed.