---
type: Policy
id: https://w3id.org/ktl/knowledge/policies/ai-covenant
title: AI covenant
description: Community norms for AI use - contributors own what they submit regardless of tooling, AI systems must not be used to autonomously post comments, replies, or messages, AI co-authorship in commit messages is discouraged, and a repository-owned agent (e.g. ktl-librarian) must commit under a bot or maintainer identity with no trailer either way and land only as a human-reviewed PR.
genre: reference
resource: AI_COVENANT.md
generated:
  by: human:noelmcloughlin
  at: "2026-10-13T00:37:18Z"
verified:
- by: human:noelmcloughlin
  at: "2026-09-10T00:00:00Z"
- by: human:noelmcloughlin
  at: "2026-09-24T01:16:00Z"
- by: human:noelmcloughlin
  at: "2026-09-26T20:48:23Z"
- by: process:ktl-librarian
  at: "2026-10-05T12:41:36Z"
- by: human:noelmcloughlin
  at: "2026-10-13T00:37:18Z"
stale_after: 2027-09-26
---

# Overview

The covenant applies to the three repositories of the project:
`knowledge-trust-ladder`, `obsidian-ktl-registrar` and
`obsidian-ktl-curator`.

The core principle is ownership: everything submitted is the contributor's,
whatever tools helped create it, and they must be able to understand, verify,
and defend it. AI review comments are suggestions, not requirements.

Discussions are for human judgement - AI systems must not be used to autonomously post comments, replies, or messages. Disclosure is required when proposing changes
to code the contributor does not fully understand, and AI co-authorship
trailers in commit messages are actively discouraged. Adapted from the LinkML
AI Covenant.

A dedicated **Repository-Owned Agent Automation** section covers the case
this repository actually runs into: an agent (`ktl-librarian`, scheduled or
interactive) proposing changes with nobody drafting alongside it. It commits
as a labeled bot identity or the invoking maintainer - never both, and never
with an AI co-authorship trailer either way; every such change lands as a PR
requiring human approval, never a direct push or an agent's own approval;
the scheduled agent runs in a read-only job with no git credentials on disk
and hands its change to a separate privileged job that runs no agent code,
which checks the "edit only `.lokf/knowledge/`" contract after the run
rather than merely requesting it; and a skill recording a person's verdict (`ktl-curator`'s `verified:
human:<id>`) may write only what that person said about that specific item, in that session,
never inferred, batched, or self-supplied.