Citations and Provenance

~22 min · citations, provenance, trust

Level 0Window Watcher

0 XP0/50 lessons0/13 achievements

0/100 XP to next level100 XP to go0% complete

Where did that fact come from?

In a long session with mixed sources — RAG hits, file reads, model priors, human messages — the origin of any given fact is the first thing that goes fuzzy. Disciplined provenance asks the model to cite its source on every load-bearing claim and refuses claims without citations on important decisions.

Citations are also eviction signals

If a fact has been cited from one source repeatedly across many turns, that source belongs in the active working set. If a fact has no citation, it is a model prior — useful for reasoning, dangerous as a load-bearing input. Treat the two differently.

Provenance versus paranoia

Asking for citations on every casual statement is exhausting. Ask on the decisions: anything that drives an edit, a commit, an authority change. Trivial conversational claims do not need provenance; load-bearing ones do.

Code

Citation contract·text

When you make a claim that affects an action:
  Cite the source: <file path>, <retrieved doc id>, <human turn N>, or 'model prior'.
When the source is 'model prior':
  State 'unverified' and ask whether to verify before acting.
When the source disagrees with another source:
  Surface the conflict; do not pick silently.

Citation-aware logging·json

{
  "claim": "auth.py uses bcrypt rounds=12",
  "sources": ["backend/auth.py:42", "docs/security.md"],
  "confidence": "verified",
  "affects_action": true
}

External links

OpenAI conversation state guide

Exercise

Take one of your AI workflows that involves edits or commits. Add a one-line citation contract to the spine. Test it: ask the model to perform a small action and watch whether citations appear.

Hint

If citations vanish under pressure, the contract is too soft. Make it a hard rule with a refusal path.