I batch-enrich a 100k+ bottling wine catalog with Claude Sonnet — style scores, identity backfill, storage potential. The safety mechanism isn’t a better prompt: it’s asking the same model the same question twice, independently. A field survives only if both passes agree (±1 for integer scores) and both report high confidence.
On edge cases the two passes disagree 10–15% of the time — and that disagreement rate is the hallucination detector. An invented grape variety almost never survives two independent runs; a real one does. Cost: 2× tokens, about $0.012 per wine — cheap insurance for data shipped to users’ phones.
The deliberate exception: the admin’s “Enrich now” button runs single-pass, because an operator watching the result is the second pass. Batch = paranoid. Interactive = fast.