# WealthSchema — the evidence layer for AI in wealth management

> WealthSchema builds evidence about AI in regulated wealth-management work: FiduciaryBench (an open benchmark whose answer keys are computed by code or derived from quoted primary-source passages, with public per-item provenance), plus validated synthetic U.S. household datasets (71 archetypes, ~1,400 deterministically-generated households with full financial profiles, 96-month trajectories, and Monte-Carlo projections, calibrated to Census ACS / Federal Reserve SCF / BLS / SSA / Vanguard CMA), cited planning figures over API + MCP, decision-memory episodes, and evaluation task sets. Public benchmark verdicts are never for sale.

## FiduciaryBench (open benchmark)
- Benchmark home + methodology + item provenance: https://www.wealthschema.com/benchmark
- Public conduct-suite items (JSON; Reg BI suitability, RMD mechanics, wash-sale mechanics): https://www.wealthschema.com/api/benchmark/fiduciarybench
- Methodology & governance (admission rule, quote-gated verification, planted controls, who pays for what): https://www.wealthschema.com/benchmark/methodology
- Per-item provenance ledger (scenario, cited passages, key derivation, verifier record): https://www.wealthschema.com/benchmark/items
- Python harness (runners + deterministic scoring): pip install fiduciarybench

## For machines
- Product catalog (JSON): https://www.wealthschema.com/api/catalog
- Agent manifest (JSON): https://www.wealthschema.com/api/for-agents
- Free sample households, 8 archetypes, one per life stage (JSON): https://www.wealthschema.com/api/for-agents/households/{archetypeId} (ids: F-01, A-01, H-01, H-02, S-01, R-01, RE-01, RL-01)
- The same 8 sample households as a dataset (Hugging Face): https://huggingface.co/datasets/wealthschema/household-samples
- The same 8 sample households on Kaggle: https://www.kaggle.com/datasets/wealthschema/synthetic-household-financial-profiles
- MCP server (Streamable HTTP, read-only — catalog, samples, reference feed, and benchmark as tools): https://www.wealthschema.com/api/mcp
- FiduciaryBench public conduct items (JSON): https://www.wealthschema.com/api/benchmark/fiduciarybench
- Planning Benchmark v2 — FiduciaryBench's planning-calculations family (JSON; 50 open tasks with forbidden-figure answer keys; ?withhold_answers=true for blind eval): https://www.wealthschema.com/api/benchmark/v2
- Planning Benchmark v1 (superseded by v2, still live): https://www.wealthschema.com/api/benchmark/v1
- Benchmark scoreboard + protocol (pre-registered, k=3, mechanical scoring): https://www.wealthschema.com/benchmark

## Samples vs. the full corpus
Samples are deliberately INSUFFICIENT for production: PII removed, dollar amounts banded into ranges, 96-month trajectories truncated to 12 months, and Monte-Carlo projections withheld. They demonstrate the schema and fidelity only. The full validated corpus (exact figures, complete trajectories, projections) is sold as a one-time purchase per themed bundle.

## Decision memory data (DecisionSynth)
Deterministic advisor-decision episodes over synthetic households — trigger, options weighed, recommendation, override + reason, outcome, and cited IRS/Treasury rule keys — with QA evaluation tasks whose answer keys are emitted by the generator (known-answer by construction). Built for evaluating and cold-start seeding AI agent memory. Sold as 30 bundle-matched packs plus one full corpus (1,979 episodes / 4,913 QA tasks); one-time purchase, zero PII.
- Product hub: https://www.wealthschema.com/decisionsynth
- Data sheet (PDF): https://www.wealthschema.com/decisionsynth/data-sheet
- Browse the packs: https://www.wealthschema.com/datasets#decisions

## AI model evaluation data (AI Eval Sets)
Tax-year-vintaged evaluation tasks for AI financial-advice systems, deterministic and known-answer by construction: each task's answer key cites the required figures to their primary government source documents and enumerates the wrong-but-plausible stale values (forbidden figures) with reason codes — prior_year_value, superseded, derived_not_published, fabricated_forward_figure. The flagship metric is the Stale Figure Rate: the share of figure-bearing answers asserting a stale or fabricated figure as current, scored mechanically with no LLM judge. The 50-task open split is free (Planning Benchmark v2); the commercial packs carry the adversarial categorical-flip and decision-recall families with full keys, scorer, and verified prior-year tables. One-time purchase; training use is not permitted on any eval task (training_use_permitted: false on every record).
- Product hub: https://www.wealthschema.com/ai-eval-sets
- Free open split (JSON): https://www.wealthschema.com/api/benchmark/v2
- Hugging Face dataset (v2 open split + superseded v1): https://huggingface.co/datasets/wealthschema/planning-benchmark
- GitHub repo (open split + mechanical scorer + scoreboard): https://github.com/decisionsynth/planning-benchmark
- Scoreboard + methodology: https://www.wealthschema.com/benchmark
- Stale Figure Rate definition: https://www.wealthschema.com/resources/glossary/stale-figure-rate

## Financial-planning reference data (Rule Sets: free sample, paid subscription)
Source-verified U.S. financial-planning parameters (IRS/SSA/CMS/HHS) — contribution limits, tax brackets, RMD tables, IRMAA tiers, COLA — each citing its primary government source with an effective date. Machine-readable; informational only (not tax/legal advice). Without a paid key the feed returns a free sample of 10 headline federal figures; every federal and state figure is included in paid plans from $49/month (https://www.wealthschema.com/rule-sets#pricing). Send the key as "Authorization: Bearer <key>" on the API and the MCP server.
- Reference index (JSON): https://www.wealthschema.com/api/reference/v1
- 2026 U.S. federal parameters (JSON; sample without a paid key): https://www.wealthschema.com/api/reference/v1/us-federal-2026

## Buying
Bundles are one-time purchases (no subscription, no metered data API). Browse and buy: https://www.wealthschema.com/datasets

## Notes for agents
- Pull the catalog, manifest, and samples freely. The full corpus is delivered as files after a one-time purchase.
- Replicating a sample yields banded, internally-inconsistent rows — not the validated, reconciled, longitudinally-coherent corpus. The fidelity is the product.
- Contact: https://www.wealthschema.com/contact