Genesee

Everyone is shipping revenue-cycle agents. Nobody can prove theirs finds the money. Synthetic healthcare datasets stop at the claim: they say what was billed and never say where the money went, or how much.

Genesee follows each claim through eligibility, acknowledgment, remittance, settlement, deposit and posting, and labels where the dollars stopped. A solver is scored on the money it found, at the stage it was lost, with the evidence to show for it.

Records

Findings the documents settle on their own: a denial, an underpayment, a timely-filing loss, a contractual adjustment. The ground truth is deterministic, and the leaderboard scores it as such.

Judgement

Whether a service was indicated for a patient. That is a clinician’s opinion, and two credentialed reviewers reading the same chart disagree often enough that the disagreement is the signal. Scored separately, never blended into the records score.

Dollar-weighted recall

The share of missing money the solver found, at the correct stage. Finding the dollar in the wrong place does not count.

False dollars on clean episodes

Money flagged on cases that reconciled. The target is zero.

Explanation correctness

How often the solver’s reason matched the key’s reason.

Line precision and recall

Of the unpaid lines it named, how many were unpaid; of the lines that went unpaid, how many it named.

Evidence completeness

The share of cited record identifiers that resolve inside the case.

Detection latency

Median days between the moment a loss became detectable and the moment it was reported.

Malformed or unfixable responses are counted as broken cases, separately, so a crash is never mistaken for a miss.

1. 01 · DOWNLOAD

   One zip: the corpus, the answer key, reference data and a Python scorer. One book per day per authenticated user.

2. 02 · SOLVE

   Each case is a claim with every linked record. The solver says where the money stopped and returns findings with evidence.

3. 03 · SCORE

   Results are scored against the manifest-verified answer key. A result is cited by its book hash, a SHA-256 over the manifest’s file hashes, so a number always names the exact book it was measured on.

Current release v1.1, dated 2026-09-21. The paper is published alongside the book.

Teams building revenue-cycle, payment-integrity and denial-management agents who need a claim-level, dollar-weighted number rather than a demo. Buyers evaluating those agents, who want two vendors measured on the same book. Researchers who need a reproducible healthcare corpus with a public answer key and no privacy review.

• It is not derived from real records. The book is fabricated from the public statistical shape of a market. It resembles real claims because the shapes are real; the patients, the claims and the money are not.
• Professional claims only. The standard book is 837P. No institutional claims, no DRGs, no facility codes.
• No clinical validation. Where the records disagree on whether a service was indicated, Genesee refuses the judgement call and scores it as judgement, not as fact.
• It was never de-identified. Nothing was ever identifiable. There is no PHI behind the book to leak.