We have LLMs now. Why are the docs still wrong?

Large language models make creating initial documentation fast, but keeping docs correct as code changes remains hard because context and audience matter. A simple code change - like a timeout moving from 30 to 60 seconds - can require different treatments across an engineering reference, a customer guide, a runbook, or an incident report. Retrieval techniques or vector indexes surface candidate passages but cannot decide whether a passage describes current defaults, a historical state, an override, or a different service. Models can produce polished rewrites that introduce incorrect options or remove important caveats, and diffs can mislead if they touch tests or conditional logic. That means evidence must come from the exact change and the live document, and reviewers need to see the old text, the proposed edit, and the supporting code context to trust a correction.

Amendary addresses this by mapping specific documentation pages to the repositories or release signals they document, then checking those pages against code changes and drafting corrections that attach the source evidence. Corrections appear in a review queue (or as a GitHub PR) so humans can inspect the precise passage, supporting diff, and rationale before merging; paid plans allow owners to auto-apply eligible edits but do not eliminate the need for ownership and judgment. The aim is fewer Slack queries and smaller, explainable documentation edits that match product changes rather than generating more noisy work.