DecisionSynth
Decisions your agents can be graded on.
An AI assistant that forgets why a client overrode policy last spring is a liability. DecisionSynth is the evaluation and seed data for that problem: deterministic advisor-decision episodes over synthetic households — trigger, options, override, outcome, and the cited IRS/Treasury figure behind each choice — with QA tasks whose answers are known by construction.
{
"episode_id": "D3-A-01-seed-103-m63",
"trigger": "promotion_or_raise",
"followed_policy": false,
"override_reason": "bracket_timing_judgment",
"outcome": "taxable_invest · approved",
"semantic_refs": [
"retirement.401k.elective_deferral.under_50"
]
}// qa.json — graded against the key, not a rubric { "task_type": "rationale_lookup", "question": "Why was policy overridden…?", "answer_key": "bracket_timing_judgment" }
Every cited figure resolves against the Rule Sets reference feed.
Households, rules, decisions.
WealthSchema builds the data layer of a synthetic wealth practice in three parts. Wealth Data Sets are the clients — 1,451 synthetic households with full financial lives. Rule Sets are the regulations — cited, current planning figures served by API and MCP. DecisionSynth is what happens where they meet: the decisions a firm makes for those clients under those rules, recorded the way an agent-memory system needs to remember them.
Six parts, none of them written by a model.
Every episode is a deterministic projection of a household's trajectory through a policy engine. No LLM writes any field — a value that can't be derived from the household, its trajectory, or a cited fact key is treated as a generator bug.
Trigger
A scheduled life event whose numbers visibly move the household's 96-month trajectory — a job loss, a tuition bill, a raise that changes the bracket math.
Options, with the rules attached
Every option a firm would consider carries a permitted flag and the cited IRS/Treasury fact keys that allow or block it — a plan loan can't be originated after separation, and the episode says why.
Recommendation
The first permitted option in the decision type's policy order. Deterministic, so two runs never disagree about what policy said.
Resolution — follow, override, escalate
A seeded behavioral draw: most clients follow, some override with a recorded reason, and overrides of $50,000 or more always escalate to compliance.
Outcome
What actually executed and with what status — approved, denied, escalated-then-approved — plus the funded amounts per source when money moved.
QA tasks, answer key included
Emitted at generation time with typed answer keys, evidence ids, and deterministic distractors. Known-answer by construction — never labeled after the fact.
Built for the teams grading agent memory — and shipping it.
Evaluate a memory pipeline on decision recall, rationale lookup, precedent search, temporal ordering, and rule attribution — scored against answers the generator knew before your system ever saw the data.
Cold-start an advisor agent with a realistic decision history: who decided what, why, what policy said, and how it turned out — without touching a production CRM.
Demo decision-audit trails and regression-test retrieval with data that carries zero PII, needs no DUA, and cites the regulatory figure behind every choice.
Ground truth you can audit, not vibes you can demo.
Deterministic, no LLM in the loop
Every field derives from the household JSON, its trajectory, its life-event schedule, or a cited fact key. Same seed, byte-identical output — drift fails the build.
Answer keys by construction
QA tasks are emitted by the generator itself, with evidence ids and distractors. Nothing is hand-labeled, so nothing is mislabeled.
Honest about what's calibrated
Households and life-event hazards are calibrated to public sources; advisor override behavior is authored, versioned, and disclosed — because no public benchmark for it exists.
Disjoint from the free benchmark
Commercial episodes generate at seeds disjoint from the public dev set and the private held-out set, asserted at build time. What you buy is never free anywhere.
The full construction methodology — including the calibrated-vs-authored table — ships with every order as a white paper, and the same document backs the data sheet.
30 bundle-matched packs. Or the whole corpus.
Each Decision Memory Pack covers the same client archetypes as its parent Wealth Data Set — episodes, QA tasks with answer keys, and the fresh households they derive from, self-contained in one download. Packs from $495.
DecisionSynth Full Decision Corpus
Every decision record in the DecisionSynth library — the full decision layer across all client archetypes: what was decided, why, what policy said, and what happened. The evaluation and seed-memory corpus for teams building or buying AI assistant memory, with zero PII.
Every pack page shows its episode and household counts before you buy.
Frequently asked questions
Is any of this real client data?+
No. Every household is synthetic and every decision is computed — a deterministic projection of that household's 96-month financial trajectory through a policy engine. No LLM writes any field, no real person appears anywhere, and there is no PII to review or agree terms over.
What exactly is in the download?+
Each pack is self-contained: episodes.json (the decision episodes), qa.json (evaluation tasks WITH answer keys), households/ (the fresh synthetic households the episodes derive from, one file per archetype, each with its 96-month trajectory), a data dictionary, and the license. Every order also includes the DecisionSynth methodology white paper as a second download.
How do the answer keys stay trustworthy?+
They're emitted by the generator at the moment each episode is created — known-answer by construction, never hand-labeled or model-generated after the fact. Generation is deterministic: the same seed and generator version produce byte-identical episodes, and every record carries its version stamps in _meta.
Are the advisor behaviors calibrated to real advisors?+
No, and we say so plainly: household demographics, finances, and life-event hazards are calibrated to public sources (SCF, NCHS, SSA, BLS), and regulatory figures cite IRS/Treasury primary sources — but override rates, reason mixes, and outcome splits are authored defaults, versioned and stamped into every episode. There is no public benchmark for advisor override behavior; claiming calibration would be false. Benchmark scores don't depend on these values — the QA tasks test retrieval of recorded decisions, not prediction of advisor behavior.
Do the pack households match my Wealth Data Set bundle's households?+
The packs cover the same client archetypes as their parent bundle, with fresh current-version synthetic households included in the pack — not the bundle's household files. Episodes, trajectories, and households inside a pack are internally consistent with each other, so the pack stands alone.
Is this the same data as the free benchmark?+
No. The commercial corpus is generated at seeds disjoint from both the public dev set and the private held-out test set — asserted at build time. What you buy is never given away, and buying it doesn't let a system imitate its way onto the benchmark scoreboard.
What does the license allow?+
One-time purchase, perpetual license for internal use — development, testing, evaluation, and demonstration within the purchasing organization. Redistribution, resale, or public re-hosting is prohibited. Full terms at wealthschema.com/license.
Stop eyeballing agent memory.
Known-answer decision episodes, from $495 · zero PII · no DUA.