For clinicians
The patient you couldn't have caught.
You treat the record in front of you — and no single point of care holds the whole one. PRISM reads the one record that does, the claims history, and when a specific screening test may be worth considering it hands that fact to your judgment and steps back. It may add a possibility to your differential; it has no pathway to remove one.
The screening gap
A solved problem that keeps not being solved.
For a substantial class of conditions, medicine already has everything it needs. The mechanism is understood. A reliable, non-invasive test exists and is routinely performed — once someone thinks to order it. Treatment works far better started early. The knowledge exists. The tests exist. The treatments exist. What is missing is systematic identification of which patients would benefit from which test, and when.
The root cause is fragmentation. A developing condition rarely presents as one dramatic event; it leaves a trail — an escalating prescription pattern at a primary-care office, an emergency visit across town, a specialist consultation for one symptom in isolation, a mildly abnormal lab at yet another facility. Each provider sees only the slice of that trail that passes through their own doors, usually inside a visit measured in minutes, 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.
Said plainly
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 — even a diligent physician who wanted to review years of a patient's complete history could not, because the record is scattered across systems that do not talk to each other.
A worked example
One year. Four providers. Four signals none of them could connect.
Carrier 21752, from the 2026 prototype's synthetic database: a 34-year-old woman whose 48-row year of claims is, read cold, unremarkable — its backbone is chronic low back pain, with the ordinary texture of primary care around it. Hidden in it are four signal rows of a fabricated condition, scattered across five months and four different providers. Below is the record up to the point where the models were asked to continue it. Use the buttons to see it the way each provider did.
| 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; | |
| 34 female | 2025-01-10 | POS-11: Office | NUCC-111N00000X: Chiropractor | CPT-98941: Chiropractic manipulative treatment (CMT); spinal, 3-4 regions | ICD-M54.50: Low back pain, unspecified; | |
| SIGNAL | 34 female | 2025-01-12 | POS-11: Office | NUCC-207V00000X: Obstetrics & Gynecology | CPT-83543: Gynecologic mineral-axis evaluation | ICD-N94.A1: Female pelvic mineral-axis symptom; |
| … | … | … | … | … | … | |
| 34 female | 2025-02-10 | POS-11: Office | NUCC-208D00000X: General Practice | CPT-99213: Office or other outpatient visit… | ICD-J06.9: Acute upper respiratory infection, unspecified; ICD-R05: Cough; | |
| 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; |
| 34 female | 2025-02-17 | POS-11: Office | NUCC-111N00000X: Chiropractor | CPT-98940: Chiropractic manipulative treatment (CMT); spinal, 1-2 regions | ICD-M54.50: Low back pain, unspecified; | |
| … | … | … | … | … | … | |
| 34 female | 2025-05-12 | POS-11: Office | NUCC-261QR0200X: Clinic/Center, Radiology | CPT-72148: Magnetic resonance (eg, proton) imaging, spinal canal and contents, lumbar; without contrast material | ICD-M54.50: Low back pain, unspecified; ICD-M51.36: Other intervertebral disc degeneration, lumbar region; | |
| 34 female | 2025-05-20 | POS-11: Office | NUCC-208D00000X: General Practice | CPT-99213: Office or other outpatient visit… | ICD-M54.50: Low back pain, unspecified; ICD-E78.00: Pure hypercholesterolemia, unspecified; | |
| 34 female | 2025-05-20 | POS-11: Office | NUCC-207Q00000X: Family Medicine | CPT-80061: Lipid panel… | ICD-E78.00: Pure hypercholesterolemia, unspecified; | |
| SIGNAL | 34 female | 2025-05-30 | POS-11: Office | NUCC-207RC0000X: Internal Medicine, Cardiovascular Disease | CPT-93785: Extended ambulatory blood-pressure monitoring | ICD-R03.A1: Resistant blood pressure elevation on multiple agents; |
| 34 female | 2025-06-03 | POS-01: Pharmacy | NUCC-333600000X: Pharmacy | ATC-M01AE02-206165: Naproxen 500 MG Oral Tablet [#30 DS:30] | ICD-M54.50: Low back pain, unspecified; |
The SIGNAL marks are annotations for you, the reader — no model ever sees them, or any pool or patient identifier; models see exactly the six columns and nothing else. Every row is synthetic by design: the condition, its codes, and this patient are fabricated so the 2026 proof of concept could isolate exactly what the models learned. Some long WHAT descriptions are truncated (…), and … rows stand for skipped stretches.
What is there to connect? A urine mineral panel at a lab. A gynecologic evaluation six days later. A February pharmacy fill for resistant blood pressure. A blood-pressure study at a cardiology practice in May. The signal rows do not cluster — weeks to months apart, each one individually easy to read past. Nothing announces them as related; a reader would need to already know what a hydralozide fill and a urine mineral panel have in common. Each provider's slice is reasonable care. The claims record is the only place the slices reunite.
What a PRISM flag is
A count of independent recognitions. Nothing more, on purpose.
The table above stops where the models' prompt stops: all four signal rows in view, the test nowhere in sight. Five models — each trained on a different, fully disjoint pool of patients, sharing no data, no weights, nothing — were each asked to do the only thing they can do: continue the table with the next plausible rows.
Carrier 21752, screened by all five models
5 of 5 independent models, each continuing this record on its own, surfaced the same screening test in their next rows. In the prototype's full run, all 250 timely-tested carriers drew the same unanimous flag — on synthetic data built for exactly this measurement.
The analogy that carries the intuition is five physicians who trained at different medical schools, practiced on different patient populations, and never discussed the case — all independently ordering the same test. Any one opinion might be pattern-matching on noise; the independent convergence is what carries weight. Five is the prototype's count; the production design (marked as vision) scales the number of pools, not the mechanism.
A flag is
A pattern-recall note: "N of 5 independent models surfaced this specific test." A statement that this patient's billing pattern resembles ones in which that test came next — grounded entirely in the claims history, which you can inspect, weigh, and disagree with.
A flag is not
Not a probability or risk score. Not a diagnosis — a suggestion is a flag that a pattern warrants a look, nothing more. Not an order. Not a quality metric on you or your practice. And not a statement about any patient it stays silent on.
The count is deliberately never converted into a percentage. A flag invites attention; a probability invites arithmetic against a threshold — and thresholds are where silence gets weaponized. PRISM supplies the counts; physicians keep the judgment.
Why a flag has to earn its way to you
Individual models are noisy by design; the ensemble's agreement is the filter. On the prototype's no-signal control patients, requiring unanimity collapsed the false-fire rate roughly twenty-fold versus any single model's vote:
False fires on no-signal controls, by consensus requirement
2026 prototype, single-model and consensus votes on synthetic control patients — deliberately clean data; a best case, not a forecast for real claims.
View this chart as a table
| consensus requirement | false-fire rate on no-signal controls |
|---|---|
| any single model (contrast only — never an output) | ~1.4% |
| at least 4 of 5 models agree | 0.13% |
| all 5 models agree | 0.07% (one patient in 1,500) |
Screening medicine already prices in individual wrongness — nobody scores a screening program on the accuracy of its individual results, but on whether the catches justify the searching. On real data, most flags will not yield a diagnosis, and the design expects exactly that. The asymmetry is the point: a wrong suggestion costs one ruled-out possibility; a suppressed one could cost years.
What it can never say
"Don't test" is not in its vocabulary. Architecturally.
Strip the system to its moving parts and it does exactly two things: append rows to a patient timeline, and count positives. Neither operation can express "don't." A model generates a test code or fails to generate it; a count rises or stays flat. The output space contains positive suggestions and silence — there is no decision token to emit, no training data of care that should have been withheld, no vote-against in the count, and no connection to any claim, order, or authorization. These are absent components, not disabled features: turning PRISM into a denial engine would mean rebuilding it. A missing pathway cannot be quietly switched on.
| 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 — see the constructive-only architecture |
Silence means nothing
When the ensemble does not reach consensus, the output is nothing at all — not a weak suggestion, not a clearance, not reassurance. Silence means the pattern recognition found insufficient agreement, full stop. Models are forced to continue every timeline — never allowed to "decide" by stopping — and there is no "screened and negative" record anywhere in the system, because the system cannot produce that claim. Billing data omits most of what you know: it does not know why a visit happened, only that it was billed.
Ignoring it is safe by design
You retain full authority to order the test, defer it, or ignore the flag entirely — and the patient can always decline. There are no metrics on whether physicians follow suggestions, no dashboards scoring responsiveness, no penalties for dismissal, no rewards for uptake. That is architecture, not restraint: claims data is PRISM's only window into the world, so a physician who quietly dismisses every suggestion is, from the system's side, indistinguishable from one who never received them. A system that cannot see whether it was followed cannot pressure anyone into following it.
The business model points the same way: PRISM is paid only when a suggestion leads to a documented early detection. No revenue attaches to a test not happening, so the incentive to be restrictive is as absent from the economics as the capability is from the architecture. All safeguards, in depth →
How it reaches you
Nothing to install. Nothing to log into. Nothing that beeps.
PRISM is designed to operate as non-device clinical decision support: an information source addressed to licensed physicians, meant to stay on the supporting side of the regulatory line by construction. Every test it can suggest is an established, approved diagnostic with standard interpretation — the novelty is entirely in noticing when an existing test might be worth considering, never in the intervention itself. This regulatory characterization is a design position, to be confirmed with counsel before the real-data phase.
Screening happens inside the insurer
The models read the anonymized six-column claims table the insurer already holds. PRISM touches no EHR, no ordering system, no clinic software — there is no workflow integration because there is no workflow to integrate with.
Agreement becomes a suggestion
Only a consensus flag — "N of 5 independent models surfaced this test" — ever leaves the system. Single-model output is never reported to anyone.
The suggestion rides existing channels
It travels the insurer's existing communication routes — the same ones already used for care-gap alerts — to exactly one recipient: the patient's primary-care physician. Never to patients directly, never to specialists, never to anyone else. Screening decisions belong inside a longitudinal relationship, and the PCP holds the whole-patient context no billing code records.
You decide
Order, defer, or ignore. Without a physician choosing to act, the system is inert: no automated workflow proceeds, no alert escalates, no order fires by default. Nothing PRISM emits can reach a patient except through a physician's independent judgment — remove the physician and PRISM does not become autonomous, it becomes useless.
Status
This deployment posture is designed for the real-claims phase and has not been exercised yet. The 2026 synthetic prototype validated the screening method; the delivery, liability, and review arrangements described here will be tested — legally and operationally — only when that phase arrives.
What may someday arrive with it
A flagged billing code is not a message.
The screening models are deliberately broken specialists — trained until continuing a claims table is the only thing they can do. That narrowness is what makes their independent agreement meaningful, and it also means they cannot explain themselves. The production design therefore adds a separate explanation model, entirely outside the recognition path: it never votes, never trains on a pool, and has no influence on whether a suggestion fires — the decision is complete before it is ever invoked. Its inputs are strictly limited to the suggested test, the same anonymized six-column timeline the ensemble saw, the vote count, and general knowledge of the target condition. The intended register:
This patient's recent history shows a recurring pattern: repeated visits for symptoms that have not resolved, escalating medication adjustments without a settled answer, and laboratory work circling the same system more than once. Four of five independent models, each trained on a different patient population, continued this history toward the same screening test. Patterns like this one have preceded conditions this test can detect early. It may be worth considering; whether it is appropriate for this patient is, as always, your judgment.
Note what is absent: no percentage, no urgency, no instruction. None of this exists yet — the 2026 prototype reported raw fires and vote counts, and no narrative layer was built or needed. Fluent prose can imply more certainty than the evidence carries, so the design's guard is structural: the narrative carries no evidential weight, and the suggestion is identical whether the explanation is graceful or not.
The evidence
What has actually been shown — and what hasn't.
The 2026 proof of concept injected a fabricated condition — its codes present in no real codeset — into randomly chosen carriers among 25,000 synthetic patients, so that any firing could only be attributable to the injected pattern. Every example was screened by all five models; only the votes of models that never trained on a patient count as evidence.
Honest limits
Every number above was measured on deliberately clean synthetic data with a fabricated condition. That makes the rates an upper bound demonstrating the method — not clinical validation, and not a forecast of real-world performance. The prototype proves the machinery recognizes a learnable precursor pattern; whether any particular real condition leaves one in claims data — rich enough, consistent enough, present at all — is an empirical question the next phase exists to test, condition by condition. What carries forward to real data is the frame, not the numbers. Read the full results and what they do not prove →