Editorial Policy

Proof-of-concept stage. The repository is private during this phase; the editorial discipline below is the standard the maintainer holds the Compendium to internally. Lint output and the operation log can be shared on request — the citation rule itself is verifiable directly from any rendered page.

This is the editorial standard the AI Safety Compendium holds itself to. It is enforced by a lint script the maintainer runs against every change before deployment.

The citation rule

Every factual claim in a concept or agenda page cites a primary source URL. The accepted source classes are:

• arXiv for papers (preprint or peer-reviewed).
• LessWrong and Alignment Forum for community technical research.
• Org blogs for first-party safety research from labs (Anthropic, DeepMind, OpenAI) and AI safety institutes (UK AISI, US AISI).
• Government and academic publications for governance, regulation, and formal analyses.

The lint script flags any uncited factual claim in a /concepts/ or /agendas/ page as an ERROR. Errors block deployment.

The only exception is explicitly-marked editorial synthesis — material under a heading such as ## Editorial note. These are clearly delimited and represent the maintainer’s judgment rather than a sourced claim.

Contradiction handling

When sources disagree on a substantive question, the Compendium presents both positions, each cited to its source. It does not silently pick a side. This applies to:

• Empirical disagreements (e.g., probability estimates of x-risk, takeoff-speed predictions).
• Methodological disagreements (e.g., whether RLHF can scale; whether interpretability can detect deception).
• Strategic disagreements (e.g., whether to support pauses; the alignment-vs-control resource allocation question).

If the disagreement is the topic of the page, both positions get equal structural treatment. If it’s incidental, the disagreement is noted with citations for further reading.

Corrections

If a page contains an error — a misattributed claim, an outdated result, a misremembered methodology — the correction process during the proof-of-concept stage is:

1. Email kevin@itforhumanity.be with the page URL, the specific claim, and a citation for the corrected source.
2. The correction is made on the affected page, the updated: timestamp is bumped, and the correction is recorded in the project’s internal operation log.
3. The maintainer replies to acknowledge the correction once the change is deployed.

The repo is private at this stage, so GitHub issues and pull requests are not accepted. Email is the single feedback channel until the project opens up.

Substantive corrections (changing a claim’s interpretation, not just fixing a typo) are also noted at the bottom of the affected page with a brief note and the date.

Provenance and timestamps

• Every page emits its updated: date in the rendered HTML header. AI engines and human readers can use this to gauge currency.
• The maintainer keeps a full git history internally; while the repo is private it is shared on request when a reader needs the audit trail behind a specific claim.
• The Compendium does not silently rewrite history; meaningful changes are recorded in the internal operation log and surfaced on-page when they change a claim’s interpretation.

Synthesis vs. claim

The Compendium distinguishes what’s known (cited claims) from how the maintainer reads the landscape (synthesis). Synthesis is allowed but must be:

• Under an explicitly-marked heading (## Editorial note or similar).
• Self-aware about its non-sourced status.
• Distinguishable from neighbouring cited material.

This matters because the Compendium is one map, not the map. Treating synthesis as fact would mislead readers about what the field actually claims.

Disclosure

• The Compendium is built with substantial LLM assistance — see methodology for the explicit auto / human-reviewed split.
• The maintainer is not a frontier-AI lab employee. The Compendium has no commercial relationship with any organisation it covers; corrections from those organisations are welcomed but receive no preferential editorial treatment.
• The maintainer’s funding sources are disclosed if relevant; the Compendium itself has no advertising and no paywalled content.

What this isn’t

The Compendium is not:

• A primary-research output. It synthesises and cites; it does not produce new empirical results.
• A peer-reviewed publication. The substitute for peer review is the citation discipline plus a transparent, on-request audit trail. Once the proof-of-concept stage ends and the repository becomes public, that trail will be one click from every page.
• A neutral arbiter of the AI risk debate. It tries to represent positions accurately, but the page-creation choices and emphasis reflect the maintainer’s judgment about what matters.
• A finished product. The Compendium is currently a proof of concept; coverage is partial, the schema is still being validated, and breaking changes to URLs or page structure are possible.

Related Pages

• about
• methodology
• index
• suggest