About the AI Safety Compendium
Status: proof of concept. The Compendium is an early-stage prototype maintained by one person. The repository is currently private, the schema and weekly workflow are still being validated, and the page coverage is partial. The rendered site is reachable to demonstrate the approach and gather feedback — not as a finished reference. Treat dates, citation density, and link-graph completeness accordingly.
The AI Safety Compendium is a weekly-updated, cross-referenced map of AI safety. It covers technical research (alignment, interpretability, evaluations, scalable oversight, control), governance and policy, risk analysis, and field-shaping commentary. Every factual claim cites a primary source.
The Compendium is published at aiforhumanity.eu. The source repository is private during the proof-of-concept stage; this page and the methodology page describe how it is built so the approach can be evaluated without repository access.
Who runs it
The Compendium is built and maintained by Kevin Huysegoms through IT for Humanity — an initiative focused on practical infrastructure for the AI safety field.
Contact: kevin@itforhumanity.be.
Why it exists
AI safety as a field is large, fast-moving, and fragmented. Important work happens across arXiv, the Alignment Forum, LessWrong, lab blogs, governance white-papers, and conference proceedings — and most readers find it through accidental discovery rather than a coherent map. The Compendium exists to be that map: a single, citable, regularly-updated reference where each concept, agenda, paper, and organisation has a permanent URL and a verifiable primary source.
The bet underneath the project is that density of citation matters more than volume of pages. A compendium that is shorter but where every claim is sourced is more useful than a longer one that drifts. The lint enforces that discipline; in the proof-of-concept stage the maintainer keeps the operation log and lint output internally and shares them on request.
What’s in scope
Broadly construed AI safety:
- Technical research — alignment, interpretability, evaluations, scalable oversight, agentic safety, control.
- Governance and policy — frontier safety frameworks, RSPs, the EU AI Act, AISI work, international coordination.
- Risk and capability analysis — empirical capability evaluations, risk decomposition, takeoff dynamics, threat modelling.
- Field-shaping commentary — work that materially shifts how the field thinks about itself.
Out of scope: pure capabilities work without a safety angle, news commentary without primary sources, and speculation without empirical or formal grounding. See suggest a source for the intake form (suggestions land in the maintainer’s private review queue while the repo is private).
How it’s built
See methodology for the operational detail. In brief: a Python ingestion pipeline fetches arXiv, RSS, and LessWrong feeds weekly; candidates are reviewed against a public manifest; approved sources are compiled into summary pages; concept and agenda pages are rewritten as the underlying landscape moves. LLM tooling assists at each step under maintainer review.
How to use it
- Read — start at index for the full catalog or overview for the field-level synthesis.
- Track changes — every page exposes a visible
updated:timestamp; the weekly sweep happens Monday 06:00 UTC. A public commit feed is not available while the repo is private. - Connect from an AI tool — see connect for the hosted MCP server (
mcp.aiforhumanity.eu) and zero-config endpoints (/llms.txt,/atlas-index.json, raw/concepts/<slug>.md). - Suggest a source — see suggest to send a paper, post, or report into the maintainer’s private review queue.
- Correct an error or send feedback — email kevin@itforhumanity.be with the page URL and what should change. While the repo is private, email is the only feedback channel; pull requests and GitHub issues are not accepted at this stage.
Editorial standards
See editorial-policy for the full set. The headline rules: every factual claim cites a primary source URL; contradictions across sources are presented with both positions, not silently resolved; pages have visible updated: timestamps; the corrections process is public.