What should LettuceDetect verify beyond RAG answers?

[adaamko]

Our current center of gravity is evidence-grounded output verification: given source material and an answer, locate unsupported, contradictory, or fabricated spans. Code answers and tool-based agent responses appear to fit that model; content safety and jailbreak detection may be adjacent rather than the same product.

Which workloads should belong in LettuceDetect? Please describe:

• the source evidence;
• the generated output;
• the failure you need localized;
• why an answer-level score is insufficient.

Concrete counterexamples and reasons a workload does not fit are especially useful.