Civic AI — 6-Pack of Care

Source for civic.ai — the bilingual (British English/Traditional Mandarin) static site for the 6-Pack of Care, a Civic AI governance framework by Audrey Tang and Caroline Green at the Oxford Institute for Ethics in AI.

Civic AI is artificial intelligence that answers to the people it affects: many small, bounded local stewards — each a Kami — that a community can own, inspect, correct, and switch off, instead of one system built to govern everyone. The 6-Pack of Care is the governance framework — six plain-language tests for AI a community can actually trust.

• The idea → start with the Manifesto; the six packs run 1 · 2 · 3 · 4 · 5 · 6, with Measures and the FAQ. The full site — illustrations, audio, case studies — is civic.ai.
• Working in the codebase → build, architecture, and conventions live in AGENTS.md.
• This README → how the repository is laid out, and how to run and contribute to it.

Built with Astro 7 and Vite+.

Quickstart

Install the global Vite+ CLI, then install project dependencies:

curl -fsSL https://vite.plus | bash
vp install

If the current shell has not refreshed its PATH, use the installed binary at ~/.vite-plus/bin/vp.

Development

vp dev     # Astro dev server → http://127.0.0.1:4321
vp build   # production build → dist/

Checks and tests

vp check
vp test

vp check runs formatting, linting, and type-aware checks. vp test is the root test suite — the renderer, root-content loader, search corpus, and static client contracts — under a 100% statement/branch/function/line coverage gate; its global setup also runs a full TypeScript compile check, the CJK-spacing gate, and the Mandarin typography gate before any test executes (this project keeps canonical content in root *.md/tw-*.md, not Astro content collections, so no astro sync step runs).

Scripts

Command	What it does
vp dev	Sync generated public assets, then start Astro with live reload at http://127.0.0.1:4321.
vp build	Production build of the static site, minified HTML, and Pagefind index → dist/.
vp check	Formatting check, linting, and type-aware checks.
vp test	Root test suite (100% coverage gate).
vp run lint	Formatting check + lint + type-aware checks + CJK spacing check + Mandarin typography check.
vp run format	Auto-fix formatting and CJK spacing.
vp run check-links	After a build, validate internal links and anchors across dist/.
vp run check:hybrid	After a build, validate the hybrid Astro/legacy-content design contract.
vp run check:astro-output	After a build, validate Astro-specific output shape.
vp run check:search	After a build, validate the search overlay, Pagefind markup, and /tw/search-index.json.
vp run check:parity	After a build, validate output parity against the historical build baseline.
vp run polis:refresh -- --commit <sha>	Refresh the checked-in Polis snapshot from one immutable upstream commit.
vp run sensemaker:extract	Print canonical narrative evidence derived from the checked-in Polis snapshot.
vp run sensemaker:regenerate	Generate each report locale twice with local Ollama and write only byte-identical canonical results.
vp run sensemaker:verify	Verify snapshot, prompt, narrative, rendered report, and manifest hashes without contacting a model.
vp run sensemaker:test	Run the isolated offline grounding and replay tests in sensemaker/.
vp run vectorize:sync-site	Upsert the Civic AI search corpus into the civic-ai-site Cloudflare Vectorize index.
vp run worker:typecheck	Type-check the isolated worker/ Cloudflare /au package.
vp run worker:test	Run the isolated worker/ route, RAG, and citation tests.
vp run en/vp run tw	Copy the canonical English/Mandarin page set to the clipboard for translation review (macOS pbcopy).
vp run import-comics	Re-import and optimise Nicky Case's comic pages into img/ (maintenance helper, not part of the build).

A Vite+ pre-commit hook runs vp staged (formatting + CJK spacing + Mandarin typography checks) on staged files, and re-validates build parity when content or build files change.

Deterministic conference replay

/conference/sensemaking/ and /conference/report/ use the same checked-in Polis snapshot. A normal build is offline: the sensemaking view derives directly from verified CSVs, while the report consumes verified bilingual narrative artifacts bound to the same snapshot ID.

Anyone can verify and rebuild both published views without Ollama:

vp run sensemaker:test
vp run sensemaker:verify
vp build

Refreshing the source or replaying the local batch-one narrative inference is an explicit maintenance operation. See sensemaker/README.md for the immutable-source workflow, current model digest, deterministic decoding settings, bilingual evidence rules, and exact regeneration commands.

Search uses Pagefind for English pages, a Fuse-backed Traditional Mandarin sidebar from /tw/search-index.json, and a separately configured worker/ /au/:question API for streamed answers — see worker/README.md for that package's own setup. The Worker retrieves from Cloudflare Vectorize binding SITE_VECTORIZE (civic-ai-site) and uses AUDREY_MODEL=nemotron-ultra with BASETEN_API_KEY plus optional CF_AIG_TOKEN to stream Nemotron Ultra through the Cloudflare AI Gateway; without those secrets it returns a deterministic excerpt/stub response for tests and local development.

Repository layout

Path	What it is
src/	Astro source: typed root-content loader, custom Markdown renderer, layouts, components, routes, search endpoints.
_data/	Global data plus the SHA-verified, immutable Polis snapshot used by both conference views.
*.md (root)	Canonical Markdown content, British English. Front-matter permalink sets the URL.
tw-*.md	Traditional Mandarin twin of each English page (served under /tw/…); kept in parity.
assets/js/	Source client scripts for the search overlay and /au answer stream, copied into generated public/.
img/, fonts/, audio	Source static assets copied through generated public/ into the build.
worker/	Isolated Cloudflare Worker for the /capacity and streaming /au/:question answer API, with its own README.
sensemaker/	Isolated local-inference, evidence, rendering, replay-manifest, verification, and test package.
scripts/	Content, validation, search, Vectorize, build, and migration helper scripts (see Scripts).
styles.css	All site styles (mobile-first; CSS custom properties).
astro.config.mjs	Astro static build config: custom-domain root, directory URLs, dist/ output.
openclaw.md, tw-openclaw.md	Human OpenClaw bootstrap guides; machine-readable endpoint is generated from _data/openclaw_bootstrap.js.
specs/	Internal design & implementation notes (unpublished).
public/	Generated Astro public directory from scripts/sync-public.mjs — never edit by hand (gitignored).
dist/	Generated build output — never edit by hand (gitignored).

Authoring content

• Each page is a single Markdown file with YAML front matter. permalink sets the URL — e.g. 1.md → /1/, manifesto.md → /manifesto/.
• Every English page foo.md has a Traditional Mandarin twin tw-foo.md served at /tw/foo/. The two link to each other through the alt_lang_url front-matter key — keep them in parity when you change either.
• Front-matter keys in use: layout, title, meta_description, summary, lang, permalink, alt_lang_url.
• Formatting rules (em dashes, four-space YAML, locked Mandarin terminology) are enforced in CI by vp check and vp test's global setup, and locally by the pre-commit hook (staged files) or vp run lint (full repo); details in AGENTS.md.

Contributing

Pull requests are welcome. By contributing, you agree to release your work under the CC0 1.0 Universal public-domain dedication (LICENSE).

Keep English (*.md) and Traditional Mandarin (tw-*.md) variants in parity, and reuse the project's locked Mandarin terminology rather than coining new translations.

Part of the Accelerator Fellowship Programme, Oxford Institute for Ethics in AI.