Our summary
An answer can name a refund status and still get that status wrong. It can also sound polite while never giving a status at all. This article treats coverage, evidence support and communication as separate checks, so the system knows whether to rewrite a claim or collect missing information.
Exact IDs and statuses can be compared in code. For indirect wording, a language-inference model asks whether the evidence supports, contradicts or says too little about a claim. A local Nimble decision model can also choose the next response, such as revising the answer or asking another question.
The author's small probe uses twelve made-up support cases, not live customer traffic. Models sometimes confuse missing evidence with contradiction, and the suggested cutoffs are not calibrated. Nimble here runs locally through Ollama; it is not a hosted Jev test. None of the labels grants permission to issue a refund.
Key takeaways
- Check presence and truth separately: an incorrect status still answers the question, while a polite non-answer may not.
- In a support check, supply evidence first and the claim second. Too little evidence differs from a contradiction.
- Let model judgments guide how prose is revised, while code retains the authoritative transaction state and permission to act.