Our summary
A reply can give a refund status and get it wrong. A polite apology may give no status at all. The article checks coverage, evidence support and communication separately, helping a system decide whether to rewrite a claim or gather information that is missing.
Code can compare exact IDs and statuses. When wording is indirect, a language-inference model checks whether evidence supports a claim, contradicts it or leaves it undecided. A local Nimble decision model can choose what to do next, such as revising the reply or asking another question.
The probe uses twelve made-up support cases, not customer traffic. Models sometimes mistake missing evidence for contradiction, and the proposed cutoffs are uncalibrated. Nimble runs locally through Ollama here; this is not a hosted Jev test. No label gives the system permission to issue a refund.
Key takeaways
- Check whether an answer is present separately from whether it is true. A wrong status still answers; a polite non-answer may not.
- Put the evidence before the claim in a support check. Evidence that says too little does not necessarily contradict the claim.
- Use model judgments to guide prose revisions. Keep the authoritative transaction state and permission to act in code.