Methodology

How I make synthetic data you can check.

This page is for the skeptic. I build one synthetic person, an archetype, then generate their whole data trail from it. Below: how the people and records are generated, how I set and hit the calibration targets, which real-world patterns the engine reproduces, how I score fidelity, and where the data still falls short of the real thing.

The generation approach

Everything starts from the archetype.

Every record renders from a synthetic person. A language model supplies the clinical knowledge, and a deterministic engine turns that knowledge into people, journeys, and records. The people come first: one synthetic person is the single source from which every data domain renders. Generation draws on the model's clinical knowledge, while rendering runs through seeded code, so the same inputs always produce the same output.

1Generative

AI generates archetypes

A large language model, grounded in published clinical knowledge and prevalence data, drafts a library of clinically-coherent patient archetypes: condition sets, comorbidity patterns, and demographics. A diabetic with CHF and stage-3 CKD reads like a real one because the model knows those conditions travel together.

2Deterministic

The engine instantiates a population

A deterministic, seeded engine instantiates those archetypes into a full synthetic population, perturbing demographics, condition burden, and a per-member utilization propensity so no two members are identical and the cohort spans well to super-utilizer.

3Longitudinal

Each member lives a 36-month journey

Every member is walked month-by-month through enrollment, condition progression, acute episodes, post-acute pathways, and persistence. High-utilizers in the first half stay high-utilizers in the second. The journey, not a per-row dice roll, decides what happens next.

4Reproducible

Journeys render any data domain

One journey is a render source, not a single file. Today it emits eligibility, medical claims, Rx fills, and revenue, with realistic coding, adjustment chains, paid-date lag, and Part D benefit phases. The same journey is what will render encounters, labs, and quality gaps. Same seed in, byte-identical output out: the rendering is fully reproducible.

The division of labor is precise. The model generates the archetypes and the shape of the journeys. Everything downstream (instantiation, the month-by-month walk, and the rendering of each data domain) runs through a seeded, deterministic engine. Given the same seed and config snapshot, the output is byte-identical, every time.

One archetype
Diabetic, CHF, stage-3 CKD · 71 · full-dual

A single synthetic person with one coherent clinical life and a 36-month journey.

Renders every data domain, all internally consistent
Eligibility & enrollmentliveMedical claimslivePharmacyliveRevenue & paymentliveEncountersliveLabs & resultsliveADT feedliveMember journeysliveQuality measuressoon

The claim that bills the CHF admit, the eGFR that tracks the CKD, the A1c that tracks the diabetes, the open quality gap, and the risk-adjusted revenue all come from the same person, so they agree by construction rather than by reconciliation.

Sampling and targeting

How I shape a population to match the targets.

A library of archetypes is the raw material. The population you receive is sampled and weighted so its aggregates land on published targets. Most of the engineering lives here, and most of it used to be invisible inside the code.

1Weighted

A weighted library

Every archetype carries a prevalence weight: how common that kind of person is. Members are drawn in proportion to those weights, so the population inherits a realistic condition mix from the first draw.

2Max-entropy

Calibration by a gentle tilt

Raw weights rarely hit every target at once. We adjust them with a maximum-entropy tilt: the smallest change to the weights that moves an aggregate such as mean risk score onto its target while preserving the rest of the shape. Newton's method solves for it in a few iterations.

3Real model

The real risk model in the loop

When we target a risk level, each archetype is scored with the full CMS HCC model: demographic score, condition score, condition interactions, hierarchies, normalization, and the V24/V28 blend. We calibrate against the real measure, so a target means what an actuary expects it to mean.

4Capture

Documentation realism

Each archetype has a coding capture rate. Its conditions are the ground truth, and only a fraction get coded onto claims in a given year, the way real charts under-document. The dataset carries both the true state and the observed record.

The calibration mindset

Every gap between our output and a target is one of two things. A definition problem means we are measuring or targeting the wrong quantity. A weighting problem means the population mix needs to shift. That framing turns calibration into a checklist instead of guesswork, and it is how every release closes the distance to the benchmarks.

Calibration · the Medicare Advantage release

Calibrated to published benchmarks, not real claims.

The sections that follow drill into our first line of business, Medicare Advantage, the same archetype engine, instantiated and calibrated for one population. The engine's aggregate outputs are tuned to match published statistics: average risk score, medical loss ratio, PMPM by service bucket, utilization per 1,000, and HCC prevalence. Every calibration target is anchored to an explicitly-cited public source and checked automatically at release.

Read this part carefully. We calibrate to published aggregate statistics. We are nottrained on real row-level claims, and no member's real record is anywhere in the pipeline. Learning patterns directly from licensed real claims is on the roadmap, and we say so plainly below.

Benchmark sourcesPublic & cited
  • CMS Advance & Final Rate Notices
    Risk-score normalization, coding-intensity adjustment, county base rates, sequestration.
  • MedPAC March Report (MA chapter)
    Utilization per 1,000, IP admits, ER visits, PCP and specialist office visits, spending mix.
  • CMS HCC Model software & coefficient tables
    V24 and V28 risk-score construction and the blended phase-in spread.
  • CMS Report to Congress (Risk Adjustment)
    HCC prevalence targets for the top conditions across the population.
  • Public MA insurer 10-Ks
    Medical loss ratio and PMPM economics (Humana, UnitedHealth, Elevance, CVS/Aetna, Centene).
  • CMS Part D Final Rule & PCUG
    Benefit-phase structure, the IRA OOP cap, LIS cost-sharing, and MMR reason codes.

From real-world pattern to eval target

Every pattern is a number I test against.

We turn published clinical and financial patterns into explicit numeric targets, each with a source and a tolerance. A release fails its audit if any headline target drifts out of band, so nothing ships on a hunch.

Real-world patternTarget we setSourceTolerance
HCC condition prevalenceDiabetes ~30%, CKD ~22%, CHF ~13%CMS Report to Congress±3 pp
Inpatient admits~275 per 1,000MedPAC March report±15%
ER visits~600 per 1,000MedPAC March report±15%
Average risk score~1.0 blendedCMS HCC model + rate notice0.90–1.10
Medical loss ratio~86%Public MA insurer 10-Ks85–92%
Cost concentrationTop 5% ≈ 50% of spendMEPS / MedPAC (5/50)shape
Clinical pathway integrityER → IP → meds → follow-upClinical guidelinessequence

Each target is checked automatically at release and printed in the report bundle with its actual value, the benchmark, the delta, and the citation. Clinical pathways are tested as sequences rather than single numbers: a heart-failure exacerbation has to flow from ER to inpatient to discharge medications to follow-up, with consistent dates and diagnoses.

Factors we model · the Medicare Advantage release

Each release captures more of the real-world detail.

The core of every version bump is factor coverage: the number of real-world phenomena the engine actually reproduces. Below is what the current v3 “Asclepius” release models for Medicare Advantage. Each factor is validated against clinical guidance or a published rate wherever one exists.

Acute episode library

v1 · ~50 pathwaysv2 · ~1,000 pathways

Coherent claim bundles for acute events (CHF exacerbation, AMI, sepsis, hip fracture, stroke) that fire stochastically and emit an ER → IP → discharge-meds → follow-up sequence with consistent dates and diagnoses.

Validated against Clinical episode pathways & published event rates.

Member-level utilization persistence

v1 · Independent per monthv2 · Sticky propensity

A latent per-member propensity multiplier, drawn once at creation, governs every utilization draw, so high-utilizers persist. Targets H1→H2 spend correlation of 0.55–0.70 and top-10% stickiness of 40–55%.

Validated against Claims-persistence literature.

5/50 cost concentration & a well cohort

v1 · Flat curvev2 · Real right-skew

A carved-out well cohort (no HCCs, minimal care) and a ~1% super-utilizer tail produce a realistic concentration curve: top 1% ≈ 25–30% of spend, top 5% ≈ 50%, and 3–6% of members with $0 medical spend over three years.

Validated against The 5/50 rule (MEPS / MedPAC).

Monthly seasonality

v1 · Uniform monthsv2 · Q1 peak, Q4 trough

Month-of-year multipliers on ER, IP, surgery, and office visits, a Q1 flu/respiratory peak (~1.10–1.15) and a December trough (~0.90–0.92), so service dates carry the winter shape real claims have.

Validated against Seasonal utilization patterns.

Dual-vs-non-dual MLR

v1 · Invertedv2 · Duals higher

A social-determinants cost multiplier on full-dual members corrects v1's inverted loss ratios. Dual MLR now runs 2–4 points above non-dual, matching real MA disclosures.

Validated against Public MA insurer segment disclosures.

HCC-driven journeys & comorbidity

v1 · Independent HCCsv2 · Correlated triads

Conditions are sampled with a pairwise correlation structure (CHF↔CKD, diabetes↔CKD, diabetes↔vascular) and distinct diagnosis pools per HCC, targeting realistic comorbidity (DM+CKD ≈ 10%, the DM+CKD+CHF triad ≈ 3–4%).

Validated against CMS Report to Congress prevalence.

Rx refill chains & Part D phases

v1 · Random cost-sharingv2 · Phase-aware

NDC-level fills with refill chains and a running per-member OOP accumulator that flips cost-sharing across deductible → initial → catastrophic phases, with LIS caps and the post-2025 IRA $2,000 cap.

Validated against CMS Part D Final Rule.

Paid-date lag & run-out

v1 · Tight ~20-day lagv2 · Full IBNR triangle

Bucket-specific paid-date development, IP and SNF lag longer than professional and Rx, producing a lag triangle smooth enough for chain-ladder IBNR, with a genuine 6–24 month run-out tail.

Validated against MA paid-development patterns.

Adjustments, denials & reversals

v1 · Coarse status setv2 · 9+ statuses, linked chains

A full claim-process enumeration (partial pays, denials, pends, adjustments, reversals, reprocessing) with every non-original row linked to its parent so adjustment chains reconcile to net plan liability.

Validated against Claim-adjudication process modeling.

The fidelity framework

We score the data the way you score a model.

Synthetic data is only worth buying if it matches reality, so we measure it like an eval. Every release runs an automated credibility audit that compares dozens of metrics against published benchmarks and rolls them into a composite fidelity score from 0 to 100.

The score is the average per-metric accuracy, where accuracy is how close each metric sits to its benchmark target. An in-band-but-off-center metric still scores below 100, so the number stays honest and does not saturate. The audit also fails the build if any headline metric drifts past its tolerance, so a regression cannot ship quietly.

Every purchased dataset carries its own full audit in the report bundle: target, actual, accuracy, and source citation for each metric, so you can re-derive the score against the exact files you received.

Example scorecardv3 “Asclepius” · 94 / 100
MetricOursBenchmarkStatus
Avg risk score1.000.90–1.10in band
Medical loss ratio86%85–92%in band
PMPM medical$983$850–1,150in band
Cost concentration (top 5%)52%~50%in band
ER visits / 1,000571550–650in band
PCP visits / 1,0005,6324,500–7,000in band
IP admits / 1,000236250–300below

Illustrative. Inpatient admits sits just under band here; we show the miss rather than hide it. Each dataset's real audit ships in its report bundle.

What the score is built from

Four dimensions, dozens of metrics.

Actuarial control totals

Risk score, medical loss ratio, PMPM by category, and utilization per 1,000, against CMS and MedPAC.

Condition prevalence

The HCC and comorbidity mix against the CMS Report to Congress, so diabetes, CKD, and CHF land at real rates.

Code-space coverage

DRG, CPT/HCPCS, revenue-code, and ICD-10 breadth against the code space real claims actually use.

Journey coherence

Every file for a member agrees: the claim, the lab, the medication, and the revenue come from one life.

Available now vs expanding

One engine, every render target, shipped honestly.

The architecture renders any data domain across any line of business from the same archetypes. We label exactly what that means in practice today, so you never plan around something that isn't live yet.

Data domains4 live
  • Eligibility & enrollment
    Member-month enrollment, demographics, plan/program, and benefit status — the spine every other domain links to.
    Available now
  • Medical claims
    Line-level institutional + professional claims: diagnoses, procedures, settings, allowed/paid, and realistic adjustment chains.
    Available now
  • Pharmacy
    NCPDP-grade drug fills with refill chains, benefit phases, formulary tiers, and net-of-rebate economics.
    Available now
  • Revenue & payment
    Payer-side revenue the way a plan receives it — capitation, risk scores, and the factors behind every dollar.
    Available now
  • Encounters
    Visit-level utilization rolled up from the journey: office, ER, inpatient, SNF, outpatient, and lab encounters with length of stay, primary diagnosis, and DRG.
    Available now
  • Labs & results
    Ordered tests with result values trended to each member's conditions. The A1c tracks the diabetic, the eGFR tracks the CKD stage.
    Available now
  • ADT feed
    HL7-style admit, discharge, and register events derived from facility and ER encounters, with patient class, facility, and discharge disposition.
    Available now
  • Member journeys
    One pseudo-chart per member: the archetype's true problem list, the HCCs actually coded this year, and a chart-note narrative. Ground truth paired with the observed record.
    Available now
  • Quality measures
    Measure-ready numerators, denominators, and gaps (HEDIS-style / Stars) rendered from each member's actual care.
    Expanding
Lines of businessMA live
  • Medicare Advantage
    Our first line, shipping today: HCC risk adjustment, MMR revenue, Part D, dual/LIS dynamics.
    Available now
  • Medicare FFS
    Traditional fee-for-service Medicare — the benchmark population behind most ACO and VBC work.
    Expanding
  • Commercial / employer
    Working-age commercial populations — different age mix, benefit design, and cost curve.
    Expanding
  • Medicaid
    Managed Medicaid with its own eligibility churn, demographics, and program structure.
    Expanding
  • ACA / exchange
    Marketplace populations with HHS-HCC risk adjustment and metal-tier benefit design.
    Expanding

Expanding render targets reuse the same archetypes and the same deterministic engine, and each ships only once it has its own calibration to published benchmarks and its own fidelity audit. None are generally available until marked live.

Roadmap

Where the fidelity goes next.

Beyond widening the render targets above, the biggest item is honest about the calibration limitation: today we calibrate to published aggregates. Next, we learn patterns directly from licensed real, row-level data.

Training on real, row-level data (v3)

Roadmap

Train on licensed real claims to derive distributions and pathways directly from observed data, moving from calibrated-to-aggregates toward learned-from-reality. No real record appears in today's pipeline; this is the planned v4 shift.

New domains: encounters, labs, quality

Roadmap

Render encounter-level utilization, ordered labs with condition-trended results, and HEDIS-style quality gaps from the same archetypes, each calibrated and audited before it goes live.

New lines of business

Roadmap

Medicare FFS, commercial, Medicaid, and ACA / exchange populations on the same reproducible engine, each with its own age mix, benefit design, risk model, and cost curve.

SNP cohort modeling

Roadmap

Explicit D-SNP, C-SNP, and I-SNP structures with their distinct utilization, revenue, and plan/contract characteristics.

Provider continuity & referral networks

Roadmap

Stable attributed PCPs and specialist panels, referral graphs, and provider-level continuity to support real VBC and network analysis.

Readmission & post-acute pathways

Roadmap

30-day readmission, SNF, and home-health pathways calibrated to clinical literature targets (CHF ~22%, AMI ~17%, ortho SNF use ~60–80%).

Don't take our word for the fidelity.

Download the 1,000-member sample, full schema, full credibility audit, no signup. Run your own checks against the numbers on this page.

Prefer the landing page first? Back to overview →