Hallucination triage battlecards
Recall the main diagnostic moves for unsupported AI claims.
Common objection "Just tell the model not to hallucinate." A customer-facing answer contained an unsupported claim. Your line We will add a refusal rule if needed, but first we need to know whether the source was missing, stale, not retrieved, or ignored. Treating hallucination as one generic wording problem. The fix must match the failed control. Source check Supported claim or plausible claim? Prompt-ops should privilege supported over plausible for operational claims. Five Whys What is the first Why after an unsupported claim? Why did the workflow allow a claim without allowed evidence? This moves the diagnosis from model personality…
Sign up free — one personalized lesson every day, matched to your role and goals.
Already have an account? Sign in