What quality checks should developers run before training on generated complaint material?

Run these checks on a generated sample before treating its labels and outcomes as training targets.

- Verify that conclusions have supporting evidence. - Check event order against causal dependencies. - Inspect labels for unsupported certainty. - Keep related case variants within one split. - Search for outcome clues in identifiers. - Review language diversity across equivalent outcomes. - Inspect grounding and disclosure risk.

The purpose is to find defects that an agent could learn as useful shortcuts. If every escalated case uses the same opening phrase, the model may learn that phrase instead of the escalation rule. If closely related variants appear in both training and evaluation, apparent generalisation may reflect familiarity with the underlying case.

Causal checks should go beyond timestamp sorting. A decision can appear after an evidence request but still depend on evidence that never arrived. The record set needs to support the reasoning attributed to the investigator.

Generated material does not automatically establish acceptable disclosure risk. Ask how grounding was translated into the release and inspect details that could retain distinctive source patterns. Record quality issues by category so corrections can address the generation process, rather than merely polishing individual examples.