How do I judge realism and disclosure risk in a generated escalation sample?

**How do I judge whether generated escalation data reflects how businesses really work?**

Ask domain staff on your side, or a trusted reviewer, to read a spread of cases and notes and mark anything that would not happen in practice: limits that are inverted, referrals that skip required levels without explanation, or notes that read like policy restatements. Disagreement is useful feedback for scoping.

**Can I test whether the output leaks anything sensitive?**

You can inspect for it, but you should not assume it is absent. The material is generated, yet generation can still produce text or identifiers that resemble real entities. Search for unusual specificity, run your own matching against names and identifiers you care about, and report anything suspicious before use. MirrorFirm does not claim that generated output is anonymous or compliant with any regime, so your own review is part of the process.

**What if the cases feel too tidy?**

Check outcome distribution and note variation. Tidy data often shows as uniform notes, evenly spaced amounts, or decisions predicted by one field. Ask for more messy cases such as missing evidence, borderline exposure and wrong-level referrals.

**Is one sample enough to decide?**

It is enough to judge structure and integrity. To judge variety, request a second sample with different seeds or a different domain and compare.

**Who decides what counts as passing?**

You do. Set the thresholds in terms of your own training or evaluation purpose.