Trust is confidence that a model or evaluation system behaves reliably enough to support decisions. It depends on evidence: an attractive offline score or a reassuring announcement is not enough if real-world behavior tells a different story. section 2.3.1 section 10.6.4
Trust can erode without an obvious drop in engagement. A movie recommender might repeatedly surface the same content, over-promote sponsored titles, or hallucinate plot summaries while click-through rate and dwell time remain healthy. Users may still feel manipulated and eventually stop returning. section 8.4.3
Build evidence for trust across evaluation and product behavior:
- Align offline metrics with the online outcomes they are intended to predict, then compare the two regularly. Offline evaluation can make an A/B test more effective; it does not replace one. section 2.3.1
- Document diagnostic assumptions and explain how metrics relate to product goals. Include underrepresented groups and edge cases rather than testing only the most common experience. section 3.5
- Instrument negative behavioral signals as imperfect proxies. Track skip streaks, abandonment after recommendation interactions, and reports as cohort-level trends; a sequence or rising category-specific rate can be more informative than one event. section 8.4.3
When offline scores look good
An agent may score well on offline task completion while production traces reveal excessive tool calls, poor escalation, or unexpected failures. Compare the offline measure with live workflow outcomes and refine it as the relationship becomes clearer. section 2.3.1
Evaluation infrastructure needs scrutiny too. For an LLM-as-a-judge pipeline, monitor prompt failures and versions, judge model behavior, cost and latency, and result drift. Re-running a fixed prompt set against a baseline can reveal changes in the judge or its API. section 9.10.4
After a failure, restore confidence by re-demonstrating reliability over time. Re-run the full validation loop against the gold-standard dataset, explain what went wrong and what changed, and temporarily increase canary and re-validation frequency. Fixing the reported case alone may leave related failures undiscovered. section 10.6.4

















