Insurance fails when one date pretends to be five
A mental-health Session may happen on Tuesday. The Claim may arrive on Wednesday. An insurer may ask for evidence on Thursday, decide it on Monday, issue remittance a week later and move the money after that. Flatten those facts into one updated date and the history becomes impossible to reproduce.
That is not an accounting detail. It affects Claims operations, complaints, reconciliation, actuarial development and any review that asks what an Organization knew at a particular cutoff.
Two times for every fact
The Heyrafiki timeline records business time and knowledge time separately. The first says when the underlying Care, decision or money fact became effective. The second says when the system received and recorded it. A valuation cutoff uses knowledge time, so a backdated adjustment first received on 2 August cannot change the view produced for 31 July.
Bitemporal modelling is established engineering, not new mathematics. The useful contribution is applying it to a complete mental-health insurance evidence chain and publishing the schema, synthetic fixture, control mapping and adversarial checks together.
The benchmark is allowed to fail
The first release has four suites. It validates the OpenAPI contract, maps every operation to an accountable capability, checks Claim and settlement identities, and reproduces an as-of outstanding amount from a seven-event synthetic history.
- Duplicate event. Rejected, because one business fact cannot appear twice under the same identity.
- Out-of-order knowledge. Rejected, because the system cannot silently rearrange what it knew and when.
- Future knowledge. Rejected, because a valuation cannot use an event recorded after its cutoff.
- Unbalanced adjudication. Rejected, because billed amount must equal payer liability, patient responsibility and adjustment.
- Settlement above liability. Rejected, because observed movement cannot exceed the adjudicated payer amount without a separately governed correction.
What an insurer or regulator can inspect
The Assurance Graph links each public API operation to a named capability, control owner, published authority source and executable evidence artifact. A reviewer can move from policy version to Claim decision, from remittance advice to independent settlement evidence, and from a public statement to the test that supports it.
The graph uses IRA Claims and market-conduct guidance, DHA certification and Health Information Exchange material, CPB licensing authority and the government API boundary as separate sources. It does not present any of them as a Heyrafiki partner or approval.
The research question starts after the controls hold
Once the history is reproducible, researchers can ask sharper questions. Which reporting delays are operational rather than clinical? How do authorization rules change access and outstanding liability? Which aggregate Benefit signals improve network planning without exposing a Person's Care? Those questions need declared cohorts, baselines, approved data, error analysis and expert review.
We are opening the benchmark for insurer, government, actuarial, research and engineering teams to add reviewed synthetic scenarios. A contributed scenario needs versioned inputs, an authority boundary, expected outputs, at least one adversarial case and a deterministic command anyone can run.
The benchmark uses synthetic data. It does not establish clinical efficacy, production throughput, regulatory approval, actuarial reserve adequacy or insurer solvency. Named collaborations are credited only after every party approves the public wording.
