The story · 2024 → next

Two years, from one page
to a working system.

PRISM began as a single page written overnight — no name, no architecture, no experiment. Two years later: a decisively positive synthetic proof of concept. This page traces the arc between, including the two mid-course corrections that taught more than the clean runs did.

The arc

A destination fixed early. Everything since has been the how.

The seed: one page, written overnight

history

PRISM began in July 2024 as a single page sent over LinkedIn. It was not written as a business plan. A conversation about a possible job at a health insurer drifted, partway through, into an idea:

The insurer was sitting on complete longitudinal claims histories for its whole population, and nobody was reading them for the one thing they are uniquely positioned to reveal — the patient who should be screened and hasn't been.

The motivation named on the page was personal, and it has not changed: missed diagnoses in the founder's own family — years of escalating treatment that a simple early test would have redirected.

The document had no name for the system; its literal title was "Health Screening Recommendations From Insurance Data Using Generative AI" — "generative AI" a gesture at a toolbox, not a mechanism. Yet most of PRISM's permanent commitments were already there, stated as plain intentions: mine the insurer's own anonymized claims data for patterns; deliver screening suggestions to primary-care physicians for review — never directly to patients, never around the physician; suggest preventive measures only, explicitly avoiding sensitive predictions; a standardized patient-profile format; publish the findings and methods.

Everything mechanical was absent. No notion yet that continuing a table could be the whole task, no ensemble, no vote — nothing that could be run, measured, or falsified. The proposal was declined, and in hindsight that mattered less than the writing of it: the page fixed the destination — anonymized claims in, screening suggestions to physicians out, care never restricted — and left every question of how unanswered.

The page is preserved as provenance, not as a specification. Where it and the current system disagree, the page is simply earlier, not wiser.

The first prototype: one bet confirmed

history

Over the following year the core vocabulary took shape on paper — disjoint pools and consensus voting, the constructive-only constraint, three-pattern learning from caught-early, caught-late, and no-signal outcomes. All of it rested on one untested bet: does next-token prediction on formatted billing sequences teach a model anything generalizable? Until October 2025, that was an argument, not an observation.

The test was deliberately minimal. Roughly 9,790 synthetic patients — about half the dataset then planned — were rendered into an early draft of the timeline format and used for one epoch of full-parameter fine-tuning on two open base models, one small and one mid-size, neither instruction-tuned. Validation was a single qualitative probe: one training patient truncated mid-history, continued by all four models — trained and untrained, both sizes.

The untrained models failed in instructive ways. The untrained small model looped, emitting the same procedure code six times in a row. The untrained mid-size model looked coherent — until a check of the full record showed it was echoing a nearly identical hospitalization from 2.5 months earlier in the same patient's history. It was pattern-matching the prompt, not modeling care. The trained models did neither: both generated novel, medically coherent futures that appeared nowhere in the patient's actual record, drawn from regularities learned across thousands of patients. The trained small model still over-applied what it learned — a breast-cancer marker ordered for a male patient — a capacity limit noted honestly at the time.

Just as telling was what the models declined to learn. The training data was left deliberately noisy — including anatomically impossible procedures — and the trained models reproduced none of it: patterns consistent across the population were learned, while errors appearing in one record were smoothed away as noise. That resilience mattered, because real claims data is never clean.

The October 2025 runs, exactly as recorded (superseded)
Historic run specs from a superseded prototype — validated only by the single qualitative probe above. Training loss is an early hint, not an evaluation result.
runbase modelhardwaredurationtokensfinal train loss
1Qwen3-0.6B-Basefour 16 GB consumer GPUs (homelab)~7.2 h~151 M0.204
2Qwen3-8B-Basetwo rented 96 GB GPUs~7.7 h~151 M0.159

This was one bet confirmed, not a system validated. There were no pools and no ensemble — both models trained on the same full dataset — no steering round, no injected condition, and no evaluation frame beyond that one hand-inspected patient. The 2026 prototype later replaced every specific of these runs. What survived is the finding itself: continued pre-training on billing sequences learns generalizable structure. Every later result stands on that.

A company, a fabricated condition, a decisive experiment

history demonstrated

In 2026, PRISM incorporated as PRISM Initiative, PBC — a Delaware Public Benefit Corporation — making the patient-outcome priority charter-level, a commitment that binds the company through funding and ownership changes.

Early that year the synthetic proof of concept began: a fictional condition, Primary Veladrin Excess, its codes present in no real codeset, injected into randomly chosen carriers among 25,000 synthetic patients across five disjoint pools — one independently trained model per pool. In a synthetic world there is nothing to diagnose, only a pattern to place and later recognize, so any firing could only be attributable to the injected signal.

Two corrections happened mid-course, and both are recorded plainly — each was more instructive than a clean run would have been:

Mid-course correction one · a data confound

The first build selected carriers hypertension-first; the models learned "sick hypertensive patient," not the injected pattern. The build was redone with uniformly random carrier selection.

Mid-course correction two · a method reversal

The planned preference-training secondary round ranked the right continuation but would not emit it in free generation; it was abandoned for pure supervised fine-tuning on the chosen rows.

By mid-2026 the full evaluation grid — every phenotype example fired against every model, 16,250 runs, zero errors — completed with a decisive result: fire when the pattern is present, silence when it is stripped, consensus collapsing the false positives.

Honest limits

The synthetic prototype proves the method works when an analogous precursor pattern exists; it deliberately does not prove that real conditions have such patterns. That is the explicit leap of faith the project now tests. The rates behind the verdict were measured on deliberately clean synthetic data — an upper bound, not a forecast. The full results, and what they do not prove →

Real claims data — the phase that has not yet begun

concept

The next phase ingests real claims data, runs the same two training rounds per pool, and evaluates retrospectively: shown only the early portion of a real patient's record, can the ensemble surface a screening the record itself only reached years later? Every historical patient whose late diagnosis is already on record becomes a test of whether the ensemble would have flagged them sooner.

The retrospective read carries weight because it mirrors how the prototype operationalized earliness — firing on a history cut before the diagnosis is the same act whether the timeline is synthetic or real.

If that holds, the first pilot follows: suggestions delivered to physicians through existing channels, with PRISM paid only when a suggestion leads to a documented early detection. As of mid-2026, none of this has begun — no real claims data has been touched, and this site says so plainly.

What stayed constant

The oldest commitments are the ones that never moved.

history concept

A recall instrument, never a predictor

The mid-2024 proposal already framed PRISM as a system that may only add a screening suggestion, and whose silence must never be readable as "no need." That constraint has shaped every build decision since — down to banning end-of-sequence at inference, so no model can ever be read as declining. How that boundary is enforced →

Agreement as the unit of evidence

Individual models were always expected to be noisy, and agreement among independently trained models was always the intended unit of evidence. Two years of building refined both ideas without weakening either — the prototype's consensus numbers are the 2024 sketch made measurable. Pools and consensus, end to end →

Who's building it

One builder, so far.

concept

Brian Jorden, founder, carried PRISM from concept to working prototype: he designed the method, generated the synthetic world, trained the models, assembled the inference fleet, and ran the evaluation — the 2026 synthetic proof of concept is his work end to end. Trusted collaborators he has shipped real systems alongside are being brought in as the project moves toward real claims data. The framing is deliberately plain: no aspirational titles, no roster ahead of the work. The team grows as the work does.

Keep going

The result, then the destination.

The evidence

The 2026 synthetic proof of concept in full: how the experiment was built, every headline number with its caveat, the ablations — and what none of it proves.

Read the evidence →

The vision

Where this goes if the real-data phase holds: production intent, explicitly marked as vision and never written as though it already exists.

Read the vision →