Questions, answered straight.

5,000 synthetic members over 36 months, in eight files that all join on member_id: eligibility, medical claims, pharmacy, revenue, labs, encounters, ADT events, and a per-member journey (a pseudo-chart with the true problem list and coded HCCs). There's also a derived/ folder with lag triangles, a PDF report with the control totals and fidelity, and the license.

Your choice at download: Parquet or CSV. Parquet is smaller and loads fast in pandas, Polars, or DuckDB. CSV opens in Excel or anything else with no extra tools. The claims files are large, which is why Parquet is the default, but CSV is there so you can look at it without installing anything.

Everything keys on member_id. A member's claims, labs, revenue, and chart all describe the same person and stay consistent, because they're generated from one underlying clinical history rather than stitched together separately.

Medicare Advantage, calibrated to a national MA population. Commercial, Medicaid, ACA, and Medicare FFS are on the roadmap, on the same engine.

The free sample is 5,000 members; the paid Complete panel is 100,000. Larger cuts are possible. The engine is seeded and scales without re-running any AI, so size is mostly a parameter.

People first. A language model writes a library of patient archetypes, which are coherent bundles of conditions, severity, and utilization. The engine samples members from that library, simulates 36 months of each one's clinical life (conditions progressing, acute episodes, medications), then renders the claims, labs, encounters, and revenue from that timeline.

No, not yet, and I'd rather be straight about that than dodge it. The patterns come from clinical knowledge plus published benchmarks (the CMS Report to Congress, the Plan Communications User Guide, and similar), not from licensed row-level real claims. Training on real data is on the roadmap. So today it's structured and calibrated like the real thing, not learned from it.

Two ways. Every dataset ships a credibility audit comparing dozens of metrics to published benchmarks, each cited, so you can check the totals yourself instead of taking my word for it. And a public-data benchmark comparison (against CMS, MEPS, and similar) is in progress, so the distributions become verifiable, not just the headline numbers. Where no public source exists, the pattern is expert-modeled and labeled as such.

Yes, fully. The dataset is a function of the libraries, the archetypes, the code, and a seed. The same seed regenerates it byte-for-byte. The AI runs once per version to build the library; generating a million members is not a million API calls, it's deterministic engine code.

HCC prevalence and risk from the CMS Report to Congress, risk scores and revenue from the MMR and Plan Communications User Guide, inpatient admits and ER visits per 1,000, medical loss ratio, actuarial value, and the 5/50 cost-concentration rule. Each metric in the audit carries its citation.

A composite of how close each metric lands to its benchmark. v3 sits at 94/100. The audit lists every metric, including any currently below band, so it's a measurement rather than a marketing number.

A few honest gaps. Inpatient admits per 1,000 run a little low right now. The lag method is volatile at early development periods, which is true of real triangles too. Procedure- and DRG-level payment-speed variation isn't modeled yet. And because it's calibrated to aggregate national benchmarks, it won't reproduce any single plan's idiosyncrasies.

Yes. Every claim carries a service (incurred) date and a paid date, with realistic, service-type-specific payment lag. This is one of the harder things to find in public data, so it was a priority.

Yes, in the derived/ folder. They're month-on-month, 36x36 cumulative paid-development triangles, overall and split by service type (inpatient, outpatient, professional, pharmacy), truncated at the evaluation date, so they're real run-off triangles with the blank lower-right you'd estimate as IBNR.

Per service type, against monthly completion curves: professional is roughly 50% then 82% of ultimate paid by development month 0 and 1, outpatient ~40%/65%, inpatient ~30%/55% with a long tail past two years, and pharmacy ~85%/94%. Each claim draws its development month from its service type's curve, so the triangles run off at different speeds the way real ones do.

Honestly, probably not. 5,000 members is thin for stable development factors. The sample is for checking the schema and the shape. For real triangle work, use the 100,000-member panel or ask for a larger cut.

Not yet. Lag varies by service bucket today (inpatient, outpatient, professional, pharmacy), not by individual DRG or procedure. DRG-level payment speed is on the to-do list.

None. It's 100% synthetic, not derived from any real person, provider, plan, or claim. There's no re-identification risk because there's nothing real underneath to re-identify.

Internal use: analytics, benchmarking, research, model development and validation, software development, testing, and demos. Not for reselling or redistribution, and not for clinical, regulatory, rate-setting, bid-pricing, or other production decisions. The full terms are on the license page.

Yes, demos are a primary use case. Build and show your product on realistic data with no PHI, then swap in the customer's real data once a contract is in place.

The 5,000-member sample is free, with all eight domains. The paid Complete panel is 100,000 members, the same shape, at $5,000.

That's where it's headed: describe the cohort, geography, or benchmark you need and the engine generates a calibrated dataset to match. For now, email what you're after and I'll scope it.

More lines of business (commercial, Medicaid, ACA, Medicare FFS), more domains (quality measures), DRG-level payment speed, a public-data benchmark comparison, custom panels on demand, and eventually patterns learned from licensed real claims.

An actuary, working mostly solo, with about a decade in and around health plans and providers. I'm keeping my name off it for now.