For health plans
You already hold the only complete record
of the patient journey.
Your claims stream — assembled for billing, standardized for adjudication — is the one place a patient's whole story exists in a single record. That is exactly what systematic early screening needs. PRISM is designed to read it inside your walls, change nothing about how care is delivered, and earn only when a documented early detection results.
The vantage point
The only assembled record there is.
Exactly one party in the healthcare system sees a patient's complete journey across every provider, every specialty, and every year of coverage: the payer. That view exists as a byproduct of billing, arrives pre-standardized, and is the entire reason PRISM is built on insurance claims.
Each provider sees their own fragment. The insurer sees all of it, because every one of those encounters generated a claim. The view is cross-provider — every physician, pharmacy, laboratory, and facility the patient touches, regardless of which health system employs them — and longitudinal — coverage persists across years, capturing the slow arc of a developing condition rather than disconnected snapshots.
Nobody designed claims data as a research asset. Its completeness follows from one blunt fact: providers who want to be paid must submit claims. And the same financial machinery enforces standardization — a rule written to make adjudication tractable accidentally produced something close to an ideal substrate for machine learning: a finite, documented vocabulary in which the same condition leaves similar traces regardless of who treated it or where.
A single provider
Sees only their own encounters. No visibility into other physicians' prescriptions, other facilities' tests, or patterns spanning specialties.
A hospital or health system
Sees only patients who choose its facilities — and loses them to outside specialists, retail pharmacies, out-of-network emergencies, and every move between cities.
EHR networks
Competing systems treat records as competitive assets. Decades of interoperability mandates have produced partial, inconsistent sharing; a record that arrives as an unsearchable document is not data.
These are not failures of effort — they are consequences of a delivery system in which patients move freely, institutions compete, and records live where care happened. The payer relationship is the one thread that follows the patient through all of it: coverage does not care which building the claim came from. The payer holds the only assembled record there is.
The honest limit of claims data
Claims describe utilization, not clinical truth. A claim records that a test was billed, not what it found; that a diagnosis code was attached, not how confident the physician was. There are no lab values, no vitals, no notes — and coding is shaped by reimbursement incentives as much as by medicine. PRISM's method is built around this constraint rather than against it: it learns patterns of utilization and suggests one constructive action — a screening test — which is a claim about what the billing record resembles, never a diagnosis.
The problem
Late detection is expensive in every currency.
For a whole class of conditions, medicine already has everything it needs — an understood mechanism, a reliable and inexpensive test, and treatment that works far better started early. Yet millions of people live for years with these conditions, because no one is systematically watching for the moment when testing would pay off. They are not undiagnosable in principle; they are undiagnosable in practice.
For the system that pays for care, waiting for complications is the most expensive possible way to discover a disease: the diagnostic test costs a tiny fraction of the emergency admissions, procedures, and chronic-damage care that follow a late diagnosis. Early detection is one of the rare places in healthcare where better outcomes and lower cost are the same intervention.
The root cause is fragmentation. A developing condition rarely presents as one dramatic event; it leaves a trail — an escalating prescription pattern at one office, an emergency visit across town, a specialist consult for one symptom in isolation, a mildly abnormal lab elsewhere — and each slice looks unremarkable on its own. The pattern is legible only across the whole journey, and the whole journey is precisely what no single point of care possesses. This is not a failing of individual clinicians. It is the predictable output of a system in which longitudinal attention is nobody's assigned role.
A real-world illustration — not PRISM evidence
Primary aldosteronism, a hormonal driver of high blood pressure, is a fair stand-in for the whole class: common among people with hard-to-control hypertension, identified with a simple blood test, effectively treated — yet by most estimates well over ninety percent of the people who have it are never diagnosed until complications develop. The condition appears here only as an illustration; PRISM's actual evidence comes from the 2026 synthetic proof of concept, not from any real disease.
Any real fix has to satisfy three requirements at once — and satisfying all three is what points back to the record you already hold:
Population-scale
Each condition is uncommon enough that you must look at everyone to find anyone. A process that waits for someone to decide a patient merits a closer look has already failed the undiagnosed majority.
Longitudinal
The evidence is a pattern spread across years of encounters, not a single alarming value. Anything built on snapshots recreates the fragmentation it is meant to repair.
Automatic
It has to run continuously in the background, over the whole population, without waiting for a trigger — the trigger not firing is the very problem being solved.
One caveat the design carries on purpose: none of this presumes every under-diagnosed condition leaves a readable trail in billing data. Whether a given condition does is an empirical question, and PRISM's condition-selection criteria treat it as one.
Zero integration
What PRISM does with data you already have.
PRISM is designed to deploy inside an insurer without changing anything the insurer or its providers already do: no new data capture, no new interfaces, no workflow changes, and no patient data leaving the building. It reads a simplified rendering of your existing claims — a six-column timeline of standardized billing codes — and nothing else. The format adds structure, not information.
| PERSON | WHEN | WHERE | WHO | WHAT | WHY | |
|---|---|---|---|---|---|---|
| SIGNAL | 34 female | 2025-01-06 | POS-81: Independent Laboratory | NUCC-291U00000X: Clinical Medical Laboratory | CPT-83541: 24-hour urine mineral-excretion panel | ICD-E87.A3: Recurrent electrolyte mineral derangement; |
| 34 female | 2025-01-06 | POS-11: Office | NUCC-208D00000X: General Practice | CPT-99213: Office or other outpatient visit… | ICD-M54.50: Low back pain, unspecified; | |
| … | … | … | … | … | … | |
| SIGNAL | 34 female | 2025-02-15 | POS-01: Pharmacy | NUCC-333600000X: Pharmacy | ATC-C03AZ08-211584: hydralozide 25 MG Oral Tablet [#30 DS:30] | ICD-R03.A1: Resistant blood pressure elevation on multiple agents; |
| … | … | … | … | … | … |
An excerpt from the 2026 prototype's demonstration patient — synthetic by design, like all prototype data. Six columns are all any PRISM model ever sees: age, sex, and coded events — anonymous by architecture, with no names, addresses, or clinical notes to leak. The SIGNAL marks are annotations for you, the reader; no model ever sees them.
No new interfaces, no workflow changes
Nobody logs into PRISM. Physicians receive no alerts from a new system, learn no new software, and document nothing extra. Suggestions travel through whatever channel you already use for provider correspondence — care-gap notices, provider portals, secure messaging — and arrive looking like communications a practice already knows how to process.
Nothing triggers PRISM
It is not a tool a physician queries or an analysis someone requests. It runs continuously in the background, working through the covered population on rolling cycles, ingesting new claims as they settle, and surfacing a suggestion only when independent models agree.
Your building, PRISM's hardware
The computation runs on hardware PRISM supplies, installs, and manages remotely, racked inside your own data center. You provide rack space, power, network connectivity, and read access to claims data; PRISM provides the hardware, all software, monitoring, and replacement units. The prototype's sizing evidence: a full-context model fits on a single 16 GB consumer graphics card, so the racks are ordinary, air-cooled, and scale linearly.
Data never leaves the building
The hardware sits on an isolated network segment inside your facility. PRISM's remote access is for managing its own machines, not for reading your data outward — the claims data stays where it already lives, and only model-management traffic crosses the boundary.
Honest limits
This deployment model is designed, not deployed. The 2026 synthetic prototype proved the method on a five-node fleet under PRISM's own control; no installation has yet run inside an insurer's facility. The open questions are the practical ones — negotiating remote-management access with a client's security team, the mechanics of the read-only claims feed, and operational cadence at real population scale. What the design fixes now is the shape of the answer: existing data, existing channels, PRISM's hardware, the client's building.
The evidence so far
A controlled proof of concept, built to be checkable.
In 2026 PRISM ran a proof of concept on an invented condition: a fictional disease was injected into randomly chosen carriers among 25,000 synthetic patients, its codes present in no real codeset — so a positive result can only mean the method worked. Five independently trained models (the prototype's ensemble; the production design envisions more) each screened every carrier and control example — 16,250 runs in all.
Agreement crushes noise: false flags on controls, by decision rule
False-flag rate on 1,500 synthetic no-signal control patients, by decision rule — the single-model figure is the rate per model run; the consensus figures are patients flagged (2 and 1 of 1,500). Measured in the 2026 prototype's full-population grid.
View the data as a table
| decision rule | false-flag rate on controls |
|---|---|
| any single model | ~1.4% |
| ≥ 4 of 5 agree | 0.13% (2 / 1,500) |
| all 5 agree | 0.07% (1 / 1,500) |
The single design choice that mattered most — voting across independently trained models — is exactly what crushes the false positives. The agreement threshold is a tunable knob, not a fixed operating point, and a flag is always reported as a count — "N of 5 independent models surfaced this test" — never as a probability of need.
Honest limits
A synthetic world with known ground truth is engineered to be provable. These numbers are the machinery's ceiling, not a real-world forecast — and the design does not need real-world numbers anywhere near this high. What the prototype proved is that the machinery works when an analogous precursor pattern exists; it did not prove that real conditions leave patterns as learnable. That is the question the retrospective phase below exists to answer, on a plan's own data. Read the full results and their limits →
How an engagement works
Retrospective first. Paid on outcomes only.
Proven on your history before it touches your present
Evaluation starts retrospectively, in silent mode, on a plan's own historical claims. Historical data already contains the answer key: patients whose late diagnoses are on record, along with the years of claims that preceded them. Run the ensemble against those histories as they looked before the diagnosis, and it either surfaces the eventual condition earlier than the record did, or it does not. No patient is contacted, no physician receives anything, and no savings are claimed until they are measured. Only results measured on your own population justify going further.
The payment chain
PRISM charges nothing for access, suggestions, or membership. It is paid only when a suggestion leads to a documented early detection — and then as a pre-agreed percentage of the savings that detection produced. Every payment traces a complete chain of evidence through ordinary claims data:
Consensus, then a suggestion
The ensemble reaches consensus on a patient and the suggestion goes to the primary-care physician through existing insurer channels.
visible in: PRISM's own batch records
The physician orders the test — or doesn't
The physician, exercising their own judgment, orders the suggested screening test. The test carries a PRISM-specific tracking code — a routine billing modifier — which also flags the screening as free to the patient: no copay, no deductible, no balance.
visible as: a claim for the test, carrying the tracking code
Diagnosis and early treatment
The result leads to diagnosis and early treatment of the condition.
visible as: follow-on claims showing early-treatment codes
Settlement from the claims trail
The pre-agreed percentage of the documented saving is paid. Because the entire trail lives in the claims stream you already process, verifying an outcome is a database query, not an audit project.
visible as: the contract's settlement, computed from the claims above
The specific mechanics here — the patient cost-sharing waiver, the tracking-code modifier, and a fee computed from documented savings — are design intent, to be finalized with counsel and plan-design review.
If the chain breaks
If the chain breaks at any point — the physician declines, the test comes back negative, no treatment follows — PRISM earns nothing and the client owes nothing. A suggestion is a flag, not a billable event. Over-suggesting doesn't pay either: a test that finds nothing completes no chain, and every suggestion must first pass a physician who keeps full authority over the order. Restraint is priced in.
A pre-agreed share of documented savings
Compensation percentages are fixed per condition before deployment, negotiated from actuarial data on the cost difference between treating that condition early and treating it late. The insurer retains the majority of every documented saving, so its net cost falls with each successful detection; PRISM's revenue exists only inside value that has already been created and measured. Fixing the percentages up front makes the arrangement modelable on both sides — you can project net savings against your own population before committing anything, and verify afterward that you kept the larger share of every saving you paid on.
Honest limits
"Documented savings" are necessarily counterfactual — the estimated cost of the late course the patient did not experience. The estimation methodology is agreed in the contract, per condition, before the first suggestion is ever sent, and is not re-litigated case by case. Projecting treatment costs from claims data is standard actuarial work for an insurer; accepting that the projection stands in for an unobservable alternative is the price of a compensation model with no other revenue source.
Why not fee-for-service
Every conventional pricing model for healthcare technology pays for volume, and volume is exactly the wrong thing to reward here. Each rejected model detaches revenue from whether patients were actually helped:
Rejected · per-suggestion fees
Revenue scales with suggestion volume — PRISM would be paid to lower consensus thresholds and flood physicians with marginal flags.
Rejected · subscription
Revenue scales with contract renewals — PRISM would be paid to optimize engagement and retention rather than detection.
Rejected · per-member per-month
Revenue scales with covered lives — PRISM would be paid identically whether suggestions work or not.
Chosen · results-based
Revenue scales with documented early detections — PRISM is paid to catch real conditions early, and only that. The entire fee is at risk: no documented early detection, no payment.
Nothing up front — deliberately
The first-client engagement is shaped so the client risks operating trust, not budget: access to its own claims history, and a place to run — power, network, and a secured space for PRISM-managed hardware. No license fee, no integration project, no per-member charge. The evidence behind the method is the 2026 synthetic proof of concept, which proved the machinery works given a clean, manufactured precursor pattern — it did not prove that real conditions leave patterns as learnable. A vendor asking a client to trust that leap should not also be asking for money up front. The retrospective's job is to replace the leap with documented outcomes; until it does, nothing is owed.
The corporate form
The charter protects you, too.
A system that reads claims histories and learns utilization patterns sits uncomfortably close to technologies that could ration or deny care — a proximity any payer's counsel will notice. PRISM is built so that repurposing is blocked twice over, by two independent locks, and a health plan that engages PRISM inherits both.
The legal lock: a PBC charter
PRISM is a Public Benefit Corporation, incorporated in Delaware in 2026. The charter names a specific public benefit — "improving patient health outcomes through earlier detection of medical conditions" — and directors are legally required to weigh that benefit alongside financial return. The binding lives in the certificate of incorporation, so it survives new investors, new management, new owners, and the departure of the founder. Anyone who acquires PRISM acquires the obligation with it. It fails only if the charter is amended by stockholder vote — a deliberate, visible act.
The technical lock: constructive-only
The deployed system counts positive suggestions only. No pathway exists to emit a denial, a risk score against a patient, or a "no need to test." Silence is never reported as clearance — it is simply the absence of a flag. There is no cost-avoidance-through-restriction product to sell, even if someone someday wanted to sell one. It fails only if the software is deliberately rebuilt.
The two locks are independent — subverting the mission would require both a formal legal amendment and an intentional re-engineering of the system, two separate acts by two separate mechanisms, each visible. No single decision, owner, or pressure defeats both at once. And the alignment runs through the revenue itself: under results-based compensation there is no version of this business that profits without patients benefiting first.
The PBC form does not, by itself, make the technology work — the synthetic prototype carries that burden, and real-data validation is still ahead. All safeguards, in depth →
Clinical decision support
What a flag looks like to a physician.
PRISM is designed to operate as non-device clinical decision support: an information source that suggests established screening tests to licensed physicians, who retain complete authority over every clinical decision. A flag is a count — "N of 5 independent models surfaced this test" — plus the timeline that produced it, reviewable by the physician. No diagnosis, no risk score, no probability. (This regulatory characterization is a design position, to be confirmed with counsel.)
4 of 5 independent models surfaced this test — a flag worth a physician's minute, never a probability. (Shown at the prototype's five-model scale.)
| action | PRISM |
|---|---|
| suggest an established screening test | yes — this is its entire output |
| order a test | no; the physician orders it, or doesn't |
| interpret results | no; PRISM never sees test results or lab values at all |
| diagnose a condition | no; a suggestion is a flag that a pattern warrants a look, nothing more |
| recommend or provide treatment | no; treatment is outside the output space entirely |
| deny, delay, or discourage care | no pathway exists — the denial pathway was never built |
One recipient: the primary-care physician
Suggestions travel through your existing channels and go to exactly one recipient — the patient's own PCP, who holds the whole-patient context no billing code records. Never to patients directly, never to specialists, never to anyone else. The physician decides. Always.
No compliance tracking — because none is possible
There are no metrics on whether physicians follow suggestions, no dashboards scoring responsiveness, no penalties for dismissal. PRISM's only window into the world is claims data, so a physician who quietly dismisses every suggestion is indistinguishable from one who never received them. A system that cannot see whether it was followed cannot pressure anyone into following it.
This posture is designed for the real-data phase and has not yet been exercised with a health plan; the 2026 synthetic prototype validated the screening method itself. The clinician's view →
Where to go deeper
The record is yours. The risk is PRISM's.
Everything above is stated with its status attached — what is demonstrated, what is designed, and what remains to be proven on real claims. The next step of diligence is the evidence itself.