A practical framework for detecting and tracking hallucinations once your agent is live.
Detection in the wild
We sample production traffic, run reference-free faithfulness checks, and alert when the rate drifts past your threshold. Hallucination is not a single failure mode, so measuring it well means watching several signals at once rather than trusting one number.
A practical framework
- Ground truth optional: score answers against retrieved context, not a golden label.
- Claim checking: break an answer into claims and verify each against sources.
- Abstention credit: reward the agent for saying "I don't know" over inventing.
- Drift alerts: track the rate over time and page when it climbs.
Closing the loop
Detection only matters if it changes behaviour. Feed flagged traces back into your guardrails and retrieval tuning, tighten the prompts that produce the most ungrounded claims, and re-measure. Over a few iterations the production hallucination rate becomes a metric you manage deliberately, not a surprise you discover from an angry customer.