Compute what should be paid — defensibly, never aggressively.
The pitch to a burned provider is not “our tech is better.” It is: the exact thing that burned you is the thing this is engineered against. Three mechanics do that work.
The engine computes what the contract allows — it never games a code to inflate a claim. A documented, contract-grounded figure survives an audit.
What was computed, from which rate, and that it was checked by a separate set of eyes — the artifact that stands up when a clawback audit comes looking.
The procedure modifier goes into the rate-match key, so a modifier-specific rate is recovered instead of abstained — more recovery, zero aggression.
The device and the return visit bill on the same code. One coding slip delays two payments. Getting the code right the first time — and catching the issue before submission — means the cascade never starts.
The hard part of billing is built — and checked by someone other than the builder.
Reconciling an 835 remittance against the contracted rate, classifying every line, and detecting the underpayment without ever minting a false number — that is the part nobody can fake. Here it is, as proof rather than a promise.
health_seam_v0/reconciliation/It compares contracted-allowed vs remitted — never against the billed charge, the mistake that makes naive tools lie — and packages the variance as signable evidence. Don’t take that on faith — reach in.
It reconciles contracted-allowed against what the payer paid, to the cent; it refuses to state a figure for a line with no rate to support it; and it never recovers more than the contracted amount — conservatism by construction. Change what the payer paid and watch the recoverable amount reproduce to the cent — found money, made defensible. It computes in front of you, to the cent, and refuses to invent a number it can’t back. This runs the reconciliation rule right here in the page, on a sample device claim. The production engine — health_seam_v0/reconciliation — is tested cent-exact against the CMS fee-schedule authority (8 / 8, $0.00 delta). Then try to make the device line invent a recovery.
| Claim line | Contracted allowed | Payer paid | Recoverable |
|---|---|---|---|
| 64999unlisted procedure — the device | — none on file | not scored | |
| 99214office visit — placement / return | $187.40 | — | |
| 95971device analysis / programming | $63.20 | — | |
| Recoverable underpayment · scored lines only | $— | ||
The device bills under 64999 — an unlisted code with no contracted rate on file. Type any number into its paid field: it stays not scored. With no rate to stand on, the engine returns null rather than mint a figure — and it never recovers more than the contracted amount, so it can’t be made aggressive. That floor is the whole difference from a system that produces a confident number because producing numbers is what it does.
Sample claim, illustrative amounts — demo-seeded, no real patient data. The arithmetic is the engine’s own; the figures are a sample you can change.
And here is what re-runs on demand — the numbers you can reproduce yourself:
64999 (the device) — no rate, so no invented number- It refuses to invent a recovery where there is no rate. On an unlisted code with no contract rate, it returns “not scored” instead of minting — a built floor a pure-generation system does not have.
- Absence is null, never a false $0. No contracted rate means the variance is null, not a phantom shortfall.
- It fails closed. A duplicate line fails the report shut rather than passing a silent wrong number; reversals and takebacks are never scored.
Independently reviewed. A separate constitutional-review seat — not the builder — read the engine and graded it pass-with-refinements: a real review that passed it while naming the work still left. The builder never certifies their own floor.
Same spec, three builders — measured against a blind estimate and a real competitor’s quote.
Each column is labelled by how we know its numbers, so the gap reads as fact, not a brag: measured on our side, an independent blind estimate on the dev-team side, a relayed quote on the competitor side.
The honest caveats, kept: our figures are commit-span wall-clock floors (parallel agents, inspection round-trips included), the blind estimate prices the entire stack, and the competitor quote is what they told us — held as stated, not independently documented. The blind estimator’s own words on agent-authored velocity: “I can’t see that from blind.”
What’s left is priced in hours — and the real date-setter isn’t code.
What’s left splits cleanly. A usable pilot is hours of code on the built core. The production wall — real PHI at network scale — runs mostly on counsel’s clock, not developer hours; and its expensive, hard-to-retrofit part, the PHI-safety engineering, is already built and tested on fixtures. We lead with the wall rather than get caught by it.
To a human-in-the-loop pilot a friendly provider could use — a few focused sessions, under a week of real building. Provider onboarding, intake, submission + appeal surfaces, one engine guard. All leveraged on the built core.
The production wall — real PHI at network scale — splits honestly into what is already engineered and what is still ahead:
The expensive-to-retrofit part is already in the code: PHI redacted by construction inbound and outbound, a single sealed egress point — six grep-able unwrap sites, three in production paths — that refuses to fire until a client grants access, standing re-verification at the moment of fire, and a “live” path that structurally cannot be constructed until it is wired. The source even flags its own not-yet-certified parts.
The honest bound: architecture built and tested on fixtures — no real patient data has flowed yet, and the egress is sealed by design until a client turns it on. Built-and-consistent-on-fixtures is not certified-against-live; this is engineering done, not a compliance claim.
The compliance floor no code can buy: signed BAAs, a HIPAA program with counsel and a risk assessment, a third-party security audit and pen-test, SRE and on-call, per-client onboarding, and formal HIPAA certification — plus the remaining engineering: encryption-at-rest and key management, audit-trail-at-scale, a certified X12 005010X279A1 serializer, authn / RBAC, and the real transport wire. Mostly the same wall on any build path, and it runs on counsel’s clock.
A built engine plus trust with operators the market burned.
The moat is not a philosophy. It is a hard thing already built, next to a kind of credibility that is hard to buy — and the honest concession that the rest is known prior art.
The resolver pulls a payer+plan’s specific contracted rate from the public Transparency-in-Coverage index without downloading the multi-terabyte bulk — so the engine reconciles against the real contracted-allowed, not the billed charge. That extraction problem is the moat; the front door is built and inspected (not yet merged to the shared trunk).
A large carrier, a large agency owner, and your own team each said the same thing inside a two-day window: they’ve tried AI in these flows and couldn’t get it dependable. A human catching drift in the loop — confirm is the only admit — answers the exact wall they hit. Honest size: three operators is a strong qualitative convergence, not a quantified market.
“Isn’t this Pega with ethical labels?” On the mechanism, partly yes. Most of the parts are prior art; we don’t claim a new algorithm. The distinctive part is narrow and real — the shape (human catches drift in the loop, not after), and a decision engine that labels its normative content as a judgment instead of laundering it into a score. We’d rather you weigh that at its true size than overweight it.
Where the first dollar comes from.
This is the one piece no surface can project, because it isn’t built from the parts above — it’s a business call. It is left here as an open slot, on purpose, for Rico to set.
First paying customer = provider underpayment-recovery — the one genuinely built thing (the reconciliation engine) pointed at the #1 provider pain. Price against recovered dollars, not seats. — the wedge (Rico to confirm): the business model, pricing, and first-customer motion are Rico’s call and are not set on this surface.
Shown as a placeholder. No pricing, revenue, or go-to-market is claimed here — this slot stays open until Rico fills it.
The instrument reconciles to the cent and won’t state what it can’t support; the review was done by eyes other than the builder; what isn’t built is disclosed, not hidden. The one open item is where the first dollar comes from — and that one is yours.
The hard part is built and independently reviewed; the rest is priced in hours; the one open question is where the first dollar comes from — and that one is yours.
The downside is disclosed, the hard part is already built, and the moat is a built engine and hard-won trust; the one open question is where the first dollar comes from — and that one is yours.