Anchor Coverage: A CI Gate for Docs a Model May Draft but Never Own

Description and use

A CI gate ensures AI‑generated documentation prose is traceable to machine‑readable facts, reducing manual review.

Problem

Generated documentation prose can contain outdated or untraceable facts, leading to human review costs that do not scale.

Outcome

The gate ensures drafted prose is traceable to machine‑readable facts, reducing human review effort and preventing stale documentation.

Workflow

• Split docs changes into Extract, Draft, Own lanes
• Extract lane pulls facts from existing sources and is script‑owned
• Draft lane allows model output only within delimited blocks
• Own lane contains human commitments and is not model‑editable
• Run anchor_gate.py against docs and claims.json
• Anchor_gate.py parses draft blocks, extracts anchors, and checks against claims
• Fail build if any draft paragraph lacks a matching claim or has too many orphans

Evidence

The gate resolves anchors against claims.json and refuses naked numbers.

Source

Split every documentation change into three lanes before any model sees the file.

Source