Decision framing: verification depth should scale with stakes and distribution. A low-stakes internal draft and a claim going out under someone's name with legal, financial, or compliance weight do not deserve the same level of scrutiny — treating every output with maximal suspicion is as much a failure of judgment as treating every output as trustworthy by default.
Signals that external validation is needed:
- Checkable numbers or citations that can be verified against a primary source
- Claims about real, named, current entities or events
- Anything time-sensitive, where "current" matters
Signals that human review is needed:
- Ambiguous judgment calls with no clean right answer
- Ethically sensitive content
- Decisions with real consequences if wrong
This is a durable judgment skill, not a fixed checklist — the calibration habit (matching verification effort to actual stakes) is what's being tested, not a memorized list of trigger words.