Where do typed hallucination spans change your workflow?

[adaamko]

LettuceDetect can return the exact unsupported or contradictory span, plus a category, instead of only a clean/flagged verdict. We want to understand when that extra localization changes an actual decision.

If you use RAG, coding agents, evaluation, or production monitoring, please share:

1. The input and evidence available to the detector.
2. What consumes the result.
3. What action a flagged span triggers.
4. Your tolerance for false positives versus missed errors.
5. An anonymized example, if possible.

This is discovery, not a feature commitment.