Synthetic Data Validation

Synthetic data validation is the difference between audit-grade datasets and toy datasets. Without strict validation, generated data accumulates errors: net-worth that doesn't match assets-minus-liabilities, FICO scores inconsistent with payment history, age inconsistent with life events, dollar amounts that don't reconcile across sections. Each error compounds when buyers depend on the data for backtesting; a single bad household corrupts batch-level analytics.

The WealthSchema validation framework operates at four levels. Arithmetic: net_worth = assets − liabilities ±$100; expense components sum to total ±$10; tax reconciliation. Internal consistency: wealth tier matches net-worth band; FICO consistent with payment history; risk tolerance matches allocation. Richness: ≥3 life events, ≥2 stress scenarios, ≥2 goals; no placeholder text; narrative fields >20 words. Archetype fidelity: income/assets/net-worth in archetype range; occupation/industry consistent; geography accurate.

Any warning at any level flips validation_passed to false — no partial credit. This was the v3 retrospective's biggest fix: the v3 prototype shipped with validation_passed: true alongside warnings, producing data that looked clean but had embedded inconsistencies. The v4 strict-fail policy enforces no-warnings-tolerated; failed households are auto-retried on Opus (more expensive but higher quality), and second failures route to manual review.

Validation runs at multiple stages of the pipeline: per-household after generation, per-bundle after overlay completion, per-corpus before catalog promotion. The cumulative validation effort is meaningful — roughly 15–20% of total pipeline cost — but is the difference between a usable synthetic dataset and one that buyers reject during pilot evaluation.