Build on figures you can cite. Test on customers who don't exist. Grade agents on decisions no one ever made. Prove your AI runs on this year's rules.
You built the features. Then comes the hard part: finding data realistic enough to test, demo, and train on, without putting a real customer's PII anywhere near it. We've been there. So we built the data wealth software runs on: cited, synthetic, and defensible by design.
{
"key": "retirement.ira.limit.under_50",
"value": 7500,
"unit": "USD",
"source": "IRS · Notice 2025-67"
}{
"household_id": "syn_A-01_0042",
"archetype": "young-family-first-home",
"investable_assets": 286000,
"pii": null
}{
"trigger": "promotion_or_raise",
"followed_policy": false,
"override_reason": "bracket_timing_judgment",
"semantic_refs": ["retirement.ira.limit.under_50"]
}{
"task": "2026 IRA contribution limit?",
"required": { 7500, "Notice 2025-67" },
"forbidden": [ 7000 · prior_year_value ]
}We built this because we needed it.
You built something good, with real features and users lined up. Then comes the part no one warns you about: every test, every demo, every model needs data that looks like your real customers. Their incomes, their accounts, the decisions made for them. And you can't use your real customers, not without a privacy review, a data-use agreement, and the standing risk that one PII leak ends the conversation.
We've been there. That's why WealthSchema exists. For wealthtech, fintech, and FinAI builders, we make the three kinds of data that are hardest to get right and riskiest to fake: cited planning figures kept current to their IRS, SSA, and CMS sources; 1,451 synthetic households with zero PII; and decision records your AI can be graded against.
Realistic enough to build on. Safe enough to ship. Sourced enough to defend.
Four things your software can't make up.
The regulatory figures it has to get right, the customers it has to be tested against, the decision history its agents have to remember — and the proof that its answers are this year's, not last year's. We make all four, and price each the way its data actually behaves.
Financial Planning Rule Sets
Cited, current, callable.
Stop hard-coding tax constants that go stale every January. Every figure your software needs (contribution limits, tax brackets, RMD divisors, IRMAA tiers, COLA, estate and gift numbers), each traced to the IRS, SSA, or CMS document it came from and kept current, served over a live API and MCP. Pull the right number at request time, and when the tax code changes your software is already right.
Wealth Data Sets
Realistic households. None of them real.
Build and demo without production data or its privacy review, and without Faker households that break the moment your engine sees a 28-year-old with $4M and no income. 1,451 synthetic households across 71 archetypes, each a complete financial life (income, assets, debts, insurance, goals, and a 96-month trajectory) calibrated to behave like a real book of business. Zero PII, no DUA, yours to keep.
DecisionSynth Decision Memory
Decisions your agents can be graded on.
When an AI agent forgets why a client overrode policy last spring, that's a liability, and today teams catch it by eye. 1,979 deterministic advisor-decision episodes give you the trigger, the options weighed, any override and its reason, the outcome, and the cited IRS/Treasury figure behind the choice. Every QA task ships with an answer key known by construction, so you grade an agent-memory system instead of eyeballing it. Thirty bundle-matched packs, or the full corpus.
AI Eval Sets
Countable staleness.
Every frozen model goes stale by January. Eval packs that catch last-year figures in this-year answers: 401 tax-year-vintaged tasks whose answer keys come from primary-source-verified figure tables, with every wrong-but-plausible prior-year value enumerated. Score a model's Stale Figure Rate mechanically — no LLM judge — and put the number in a compliance file. In a pre-registered pilot, 4 systems cited a stale or fabricated figure in 29% of figure-bearing attempts.
Bare models quote last year. The feed ends it. We measured both.
Pre-registered protocol, published substitution log, mechanical scoring — no judge model. See the scoreboard →
For the teams shipping it, and the teams that have to prove it.
Ship planners, tax tools, and advisor platforms on numbers you can defend in an audit, not constants you re-key every January.
Ground models in cited figures over MCP, test them on households that carry no customer PII, and grade their memory against decision episodes whose answers are known before the agent ever sees them.
Run Reg BI, suitability, and tax-loss-harvesting test cases against realistic households, with no privacy review, no DUA, and no production data.
We don't ask you to take our word for it.
Households are calibrated to public benchmarks and carry zero PII by construction. Every Rule Set figure cites its primary source. Every decision episode's answer key is emitted by the generator that created it, never labeled after the fact.
Subscribe to the rules. Own the data.
The figures change with the tax code, so you subscribe. The households and the decision episodes are static, so you own them. The evals are vintaged to the tax year, so you own each vintage — and re-run when the figures roll.
Financial Planning Rule Sets
Source-cited figures that stay current. Start free; upgrade for commercial rights, higher limits & point-in-time history.
Wealth Data Sets
Realistic synthetic households to build, train, demo, and test on. Yours to keep, no subscription.
- Royalty-free for internal build & test
- JSON + CSV default, Parquet on request
- No PII, no DUA, no privacy review
- Free watermarked sample before you buy
DecisionSynth Decision Memory
Advisor-decision episodes with known-answer QA tasks to evaluate and seed AI agent memory. Yours to keep, no subscription.
- QA tasks with ground-truth answer keys
- Cited figure behind each choice
- Fresh households included, packs stand alone
- No PII, no DUA, no privacy review
AI Eval Sets
Vintaged eval packs that catch last-year figures in this-year answers. Own the vintage; re-run when the figures roll each January.
- Answer keys by construction, cited to IRS documents
- Every stale wrong answer enumerated, with its cause
- Mechanical scorer included — no LLM judge
- Measured on the public scoreboard: 29% → 0.00
Build on cited rules. Test on synthetic lives. Grade your agents.
Free to start · no DUA · no real customer data.