The platform
A claim arrives, is read into one consistent view, is edited across twelve layers, is priced in integer cents, and leaves with a numbered trace attached. Everything after that — review, appeal, simulation, reporting — reads that record rather than re-deriving it.
Intake
Scan, validate the envelope, split, evaluate, reconcile, record. Six channels run the same path, so a claim cannot arrive by a route that skips a control.
Which plan a delivery belongs to is decided by its credential — never by a value inside the file. An unrecognised sender is refused rather than accepted on a best guess.
The execution trace
Every rule the engine considered appears in order — matched or not — with the conditions it evaluated, the values it read, and why it was skipped when it was. An appeal argues with this, and usually loses.
A claim is never edited in place. A re-evaluation keeps the prior decision beside the new one, so the history of what the platform recommended is itself a record you can read.
Money
Integer cents, a frozen clock and a fixed line order. Deductible by benefit class, coinsurance, annual maximum, a separately tracked orthodontic lifetime maximum, waiting periods, age limits, exclusions and alternate benefits.
An alternate benefit changes the allowance only. A crown paid at a filling's allowance keeps major-service coinsurance and its waiting period — because the service the dentist performed did not change.
Every total is also the exact sum of its per-line parts, including after coordination of benefits.
In network, the excess over the maximum allowable is a contractual write-off the member never owes. Out of network there is no contract, so it is member exposure — and coordination never reduces it.
Applied per line by allowed amount. A claim naming another carrier as primary with no payment stated is not priced at all — $0 from the primary is exactly what produces a full and wrong payer share.
A recommended correction that never overwrites the adjudicator. A right total with a wrong composition is a discrepancy, and relabelling a denial as a write-off is one too.
Contractual, alternate-benefit and edit-driven savings with no overlap, summing exactly to what pricing wrote off. Cost sharing sits under "not savings" so nothing is mistaken for one.
Rule governance
Your policy team authors in a no-code builder or the rule DSL. The road from draft to active is the same for a rule a person wrote and a rule a model drafted.
Baseline and candidate evaluate under identical context — frozen clock, the same history, the persisted reference snapshot — so every reported change is attributable to the candidate rule and nothing else. New rules also simulate against paid history, and the money arm fails closed rather than counting an estimate as a charge at risk.
Overturns flow back into each rule's measured precision and into the baselines new rules are tested against. A rule nobody has appealed reports UNMEASURED — never a flattering 0% overturn rate.
An exposure-ranked queue with SLA. Overrides are write-once and audited, need a reason and a reviewer role, and high-dollar overrides need an approver. Edits marked binding cannot be overridden at all — authors can never resolve their own rule's output.
FWA analytics
A reproducible relationship view — providers at one practice address, the providers one patient has seen, bounded multi-step connections — each finding carrying the claims behind it. No score, no rank, no risk language, and no path from this view into a decision.
Indicators run a shadow → measured → active lifecycle, and promotion re-reads the cited precision measurement inside the commit. Two indicator libraries are kept deliberately unmerged with their overlap declared.
Enterprise ecosystem
The 835, 277 and X12 loop closed end to end. Licensed content packs with a Content Update Manager, and administration separated from claims data so authored content and member information are never in the same custody.
A pack rule arrives as a draft, not as production policy. A deployment that cannot verify content says so rather than trusting it — and a pack can never activate itself.
Performance
A benchmark without its conditions is a marketing number. Everything below was measured on a single developer-class machine — one process against a local database, not payer-scale infrastructure. Treat it as evidence about the shape of the engine, not as a capacity promise for your environment; we run the benchmark on your hardware during evaluation.
Forty-five minutes with the engineers who built the engine. Bring the edit your team disagrees on internally — that is the one worth watching.