Policy Q&A tool: handoff notes for whoever owns 03-governance

What it does

Takes a plain-language question, searches the actual markdown files in 03-governance/, and returns the top matching sections with the source file, heading, and an excerpt, plus a fixed disclaimer that this is a citation finder, not legal advice. It never answers from a model’s memory of what governance policy “probably” says; it only returns things that are actually written down in the corpus.

Why lexical search instead of embeddings

This corpus is small (a dozen or so markdown files) and changes by hand-edit, not by bulk ingestion. A keyword-overlap search is exact, inspectable, and has zero infrastructure (no vector DB, no embedding model, no index to keep in sync). If the corpus grows past a few hundred documents, revisit this, but don’t reach for embeddings before you’ve outgrown what a simple search can do, that’s added complexity with a real maintenance cost.

The one invariant that matters

If nothing in the corpus actually addresses the question, the tool returns an empty result set (no_match: true), never a fabricated-sounding answer. That’s tested (test_no_fabrication_on_irrelevant_question). If you’re extending this and tempted to add “if no match, summarize what we might guess,” don’t, that’s the exact failure mode this tool exists to avoid.

How to keep it accurate as governance evolves

This tool reads 03-governance/*.md directly, live, every time it’s called. There is no separate copy to keep in sync, no cache to invalidate. Edit a governance doc, the next question against it reflects the edit immediately. That’s a deliberate design choice: a policy tool that can drift out of sync with the actual policy is worse than no tool.

If something looks wrong

• It’s not finding an obviously relevant section: check whether that section uses different wording than the question. This is keyword overlap, not semantic search, “comp” and “compensation” are different tokens to it. Either rephrase the question or, if it’s a common phrasing gap, that’s a signal the doc itself could use clearer headings.
• You want it to also search 01-use-cases or the prompt library: change GOVERNANCE_DIRNAME to a list and extend load_governance_docs, that’s a small, safe change. Think first about whether mixing governance citations with prompt-library content in one search makes the results more or less trustworthy, that’s a judgment call worth making deliberately, not defaulting into.

Where this plugs into the rest of the playbook

This tool’s disclaimer language mirrors the “not legal advice” framing already used throughout 03-governance/ai-use-policy.md and 03-governance/quick-reference-checklist.md. Keep that consistent if you edit either.