sgit newsroom v0.1.29 · snapshot 2026-09-24

Reading room · riskmandate.ai

On this page

Reading room / riskmandate.ai · raw text · live ↗

From riskmandate.ai, the page as fetched on 2026-09-25 · open the live page ↗Everything on this sheet is the source site's own text; the newsroom's chrome is outside it.

riskmandate.ai — read this first

Rendered from CLAUDE.md in the repository. The text below is that file. Source: https://riskmandate.ai/admin/agents/claude-md/↗ · noindex · written by scripts/site/build-admin.mjs

This is the public site for RiskMandate, served byte for byte from site/ by GitHub Pages. No build step, no framework. The site sells the Agent Behaviour Policy (ABP): for one agent in one deployment, everything it can do, what it was authorised to do, the gap, and what stands in the way. Fifteen template policies exist as encrypted vaults, read live on agent-behaviour-policy.html. The Insurability Index is where this goes; we are on the first step.

Do not re-read the repository to learn it. .claude/onboarding/ is the condensed version of everything a previous agent had to read in full. Start at .claude/onboarding/00-start-here.md and follow the reading order for your task. Budget: ten minutes, not an hour.

The rules that are not optional

These hold for every page, every vault, every brief and every commit. Breaking one is not a style problem; it is a correctness problem, and the site's own tests catch some of them.

  1. No score on a behaviour policy. No rating, level, traffic light or verdict, anywhere, including in data. The Index scores the deployment; the ABP never scores anything.
  2. Never ADP. The acronym is ABP, spelled out as Agent Behaviour Policy at first use.
  3. Never "the policy" alone. Say the ABP or the behaviour policy. On this site policy is also the insurance instrument in the Licence to Operate demo.
  4. Never test somebody else's system. Read the vendor's pages, quote them, date them. Where their pages disagree, publish the contradiction unresolved. A row becomes measured only on a system we are entitled to run.
  5. No verdict on a named third party. Facts, dates, sources; no evaluative adjective.
  6. No conformity language. Not certified, compliant, conformant, accredited, aligned.
  7. Never reproduce a standards body's text. Titles only. The EU regulation may be quoted.
  8. Do not manufacture assurance. A prohibition carries its barrier; a number carries its provenance; a claim carries its date.
  9. British English. Behaviour, authorise, licence (noun). Never rung: the lead finds it an odd word, so the ladder has steps. The site's voice is plain, sourced and unhurried; it says what was not done beside what was.
  10. No write credential in site/, ever. Read keys are public by design and printed on purpose. A test fails the build if anything write-shaped lands in the deployed tree.

The mechanics, in one screen

bash scripts/run-locally__riskmandate_ai.sh           # http://localhost:10070/  (localhost, not 127.0.0.1)
node scripts/site/generate.mjs                        # markdown twins, sitemap, llms.txt, 404 — rerun after any page edit
node scripts/site/build-admin.mjs                     # the admin console, site/admin/ — rerun after touching docs/, .claude/, a register
npm run check                                         # everything CI runs: tests + every --check
node scripts/site/release.mjs 1.17.0 "Title"          # cut a version: notes stub + restamp every page

Working alongside other agents

Other agents are usually working on sibling claude/* branches at the same time, all merging into dev. The rules are in .claude/onboarding/04-rules-of-engagement.md; the short form:

Where things are

You needGo to
the shortest path into a task.claude/onboarding/00-start-here.md
every doc, brief, page family, script and register, one line each.claude/onboarding/01-map.md
the ABP model without reading five briefs.claude/onboarding/02-abp-model.md
where we are and what is next.claude/onboarding/03-state-and-next.md
how to add a page, a vault, a release, an edition, a brief.claude/onboarding/05-workflows.md
a task ready to pick up.claude/briefs/
a prompt to start a common job.claude/commands/ (also slash commands in Claude Code)
the same index, on the sitesite/admin/ — the console: what needs the lead, the board, every brief as a page

Write for the reader who arrives after you. Commit messages here say what changed and why in the site's own voice; look at git log for the register. No model identifiers in anything committed.