Skip to content

AI covenant

Policyreferencepolicies/ai-covenant

Nobody has checked this yet

still a draft

Written 2026-10-28 by ktl-librarian (automation).

The full record
Checks
none recorded
Sources
References
Graph
Open this concept in the knowledge graph
IRI
https://w3id.org/lmodel/github-actions/knowledge/policies/ai-covenant
Licence
CC BY 4.0
Files
MarkdownJSON-LDSource on GitHubHistory on GitHub

AI_COVENANT.md sets the norms for AI-assisted work in the repository.1 Its core principle is that everything a contributor submits is theirs, whatever tools helped. The contributor must understand it, verify it, explain it in review and make it meet the project’s standards.

The covenant says that the repository models someone else’s platform, which makes one failure mode worse here than elsewhere: a field, value or rule that a model produced, presented as what GitHub does. So:

  • a schema description says what its source says. Every class, slot and enum value renders a page of GitHub’s documentation, SchemaStore’s schema at the pinned commit, or GitHub’s SLSA build type, and its class or enum names that page in see_also;
  • GitHub’s documentation is cited, never copied, and the schema header records the date its pages were read;
  • the SchemaStore copy is upstream text: never edit it, replace it from the commit that NOTICE records, and upstream-watch.yaml checks that it still matches;
  • everything the schema defines lives under gha:, and GitHub’s own names, such as event names and permission scopes, appear only as values the schema documents.

AI review comments are automated quality checks, not human reviews, and a pull request still needs a person’s approval. The scheduled agent that Knowledge Librarian runs commits as knowledge-librarian[bot] or as the maintainer who ran it interactively, never both, and its commits carry no AI co-authorship trailer. Every change it proposes arrives as a pull request that a maintainer approves. It runs in a read-only job with no git credentials on disk and hands its change to a privileged job that runs no agent code, and its contract to edit only .lokf/knowledge/ is checked after it runs. A skill that records a person’s judgment writes only what that person said about that item, in that session.

Disclosure is required for a fix to code that the contributor does not fully understand, and for a schema description that an AI drafted, together with the page it rests on. Routine AI use needs no disclosure, and AI co-authorship in commit messages is actively discouraged. The covenant says that it is adapted from the Knowledge Trust Ladder AI Covenant, which is adapted from LinkML’s.

  1. AI_COVENANT.md, every section ↩