How synthetic households are built and validated
Synthetic household data is the foundation behind every Wealth Data Set on this site. This page is the public methodology index — written for compliance teams evaluating the data for examiner use, for engineering teams integrating the JSON, and for academic researchers citing the corpus.
The corpus holds itself to a strict internal-consistency contract: if a generated record disagrees with itself in any way — arithmetic, schema, narrative — it is rejected and regenerated. The result is a dataset your data team can trust without spot-checking.
Six commitments behind every record
These commitments are codified in the generation pipeline and tested on every refresh. They are the answer to the only question that matters when buyers evaluate synthetic data: can I trust this?
Schema-first generation
Every household is generated from a single canonical Zod schema. If a record fails the schema, it is discarded and re-generated — invalid data never touches disk.
Consistency contract
All downstream artifacts — overlays, longitudinal trajectories, tax calculations — are pure projections of the canonical household JSON. No secondary math, no renderer drift.
Strict validation gate
Two-pass validation: deterministic checks for arithmetic and schema, then LLM-assisted review for narrative coherence. Any warning fails the household — no soft passes.
Field-level provenance
Every field carries documented type, range, and derivation logic. Methodology PDFs ship with every Data Set so your data team can audit any number end-to-end.
Refreshes with transparent diffs
Tax law changes, market shifts, and new archetypes flow through the same pipeline. Refreshed corpus versions ship with a changelog showing exactly what moved and why. Cadence and pricing are still being defined.
Synthetic by construction
No real individuals, no GDPR exposure, no data use agreements. Sensitive overlays (race/ethnicity, religion) appear only on the bundles that explicitly require them.
Methodology documents
Long-form references covering the full generation pipeline, validation logic, and refresh process. Every bundle ships with a per-bundle Methodology PDF documenting field derivations specific to that Data Set.
How Synthetic Households Are Generated, Validated, and Refreshed
Every household in the catalog passes through the same pipeline before it ships: archetype assignment, overlay application, a strict consistency check, and an LLM-as-judge pass that can still reject a record after all that. Here's exactly how it works, stage by stage.
The Planning Benchmark Pilot of Record: Protocol, Scoring, and Results
The pre-registered protocol behind the Planning Benchmark scoreboard: 40 tasks, 4 systems from 3 labs, k=3 attempts, mechanical scoring against primary-source-verified answer keys — including the decision rule, the substitution log, and the standing caveats.
Want the per-bundle methodology PDF?
Every Wealth Data Set ships with a Methodology PDF describing the field derivations, eligibility rules, and statistical calibration for that specific bundle. Browse the catalog to see which bundle fits your use case.