Cooperative alignment

A working essay about minds, values, and the possibility of cooperation between powerful AIs and humans, with a short version and an interactive argument map.

• The reading edition presents the full essay in fourteen chapters.
• The focused map explains one claim at a time, with supporting claims, shared premises, and links to the essay.
• The whole-map overview shows every claim and connection for reviewing the structure of the argument.

The site is static HTML, CSS, and JavaScript. It uses no external fonts, analytics, JavaScript packages, or hosted assets. The Python build uses only the standard library; the production container serves the generated files with Caddy.

Build and preview

Use Python 3.10 or later:

python3 build.py
python3 check_build.py
python3 -m http.server 8000 --bind 127.0.0.1 --directory public

Open http://127.0.0.1:8000/ for the essay, /map.html for the focused map, or /overview.html for the complete graph. Links to individual claims use fragments such as /map.html#claim-H3.

public/ and build-manifest.json are generated and excluded from version control. The build validates graph references, reachability, and dependency cycles. check_build.py checks the rendered essay against its source, map payloads, local links and anchors, and the complete set of published assets.

The default URL base is /. To build for a static host mounted at a subpath:

BASE_PATH=/cooperative-alignment/ python3 build.py
python3 check_build.py

The equivalent --base-path /cooperative-alignment/ flag overrides the environment variable. Configure the host to serve the generated directory at that path. Run python3 build.py again to return to a root-path build.

Edit the argument

• content/essay.md is the full essay source. Its paragraphs are preserved exactly in the reading edition and the downloadable Markdown file.
• short-version.json contains the short version and its chapter references.
• argument-map.json contains every claim, explanation, essay reference, and dependency. Shared premises reuse a claim ID.
• build.py defines chapter boundaries by their opening paragraphs. Update those markers if editing a chapter's opening text.
• map.template.html, overview.template.html, and build_map.py render the two map views, including complete text fallbacks when JavaScript is disabled.

Each map edge names a target claim and one of five relations: supports, requires, explains, illustrates, or qualifies. Keep IDs stable so existing claim links continue to work. Rebuild after content changes, then review both map views and the linked essay passages.

Deploy on Railway

Create a Railway service from this GitHub repository. Railway detects the root Dockerfile and builds the site automatically. Set the deployment healthcheck path to /healthz, then generate a public domain in the service's networking settings. No volume or application secrets are required.

The multi-stage image builds at / using Python and copies only generated assets into the Caddy runtime. Caddy listens on Railway's PORT environment variable (8080 by default); Railway handles public HTTPS. The /healthz endpoint returns HTTP 200. Missing files return HTTP 404.

See Railway's Dockerfile documentation and healthcheck documentation.

To build and try the same image locally:

docker build -t cooperative-alignment .
docker run --rm -p 8080:8080 -e PORT=8080 cooperative-alignment

Then open http://127.0.0.1:8080/. Raw transcripts, QA artifacts, and local runtime files are not included in the image; the essay's downloadable Markdown is an intentional public asset.

Optional local macOS service

install.py maintains the existing local background-service workflow. It builds explicitly for /cooperative-alignment/, copies generated assets into the user's Application Support directory, and installs a LaunchAgent that serves them on loopback port 18764. Tailscale routing is configured separately. This workflow is independent of Railway and is not run by the Docker build.