The 2024 Seed: A One-Page Proposal
Status: history — the 2024 seed, preserved as provenance.
PRISM began in July 2024 as a single page written overnight and sent over LinkedIn — no name, no architecture, no experiment, but already carrying most of the commitments that still define the project.
Where it came from
The page 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. That evening the idea became a one-page proposal, drafted overnight and sent back the next day.
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. That experience is the same one that later shaped the choice of demonstration problem; the class of condition it points at is described in the screening gap.
The document had no name for the system — its literal title was "Health Screening Recommendations From Insurance Data Using Generative AI." "Generative AI" was a gesture at a toolbox, not a mechanism; the page contained no model design of any kind.
What the page already carried
Read today, the striking thing is how many of PRISM's permanent commitments were on that first page, stated as plain intentions.
| commitment as written in 2024 | where it lives now |
|---|---|
| Mine the insurer's own anonymized claims data for patterns | the insurance vantage point, anonymous by architecture |
| Deliver screening suggestions to primary-care physicians for review — never directly to patients, never around the physician | clinical decision support |
| Suggest preventive measures only; explicitly avoid sensitive predictions such as life expectancy | constructive-only, recall, not prediction |
| A standardized patient-profile format so that diverse datasets train and evaluate consistently | the patient timeline format |
| Publish findings and methods; explore an open, public version beyond one company | research and publication, open collaboration |
Two of these deserve a note. The instruction to avoid sensitive predictions and focus solely on preventive measures is the earliest ancestor of the constructive-only guarantee — a boundary asserted years before there was any architecture to enforce it. And the "universally applicable patient profile template" is the intuition that eventually hardened into the six-column timeline: the page already understood that the format is what lets diverse data sources train one kind of model, even though it could not yet say what that format was.
What it lacked
Just as instructive is everything the page did not contain. None of the machinery existed yet — not even as a sketch. There was no notion that continuing a table could be the whole task, no ensemble, no vote, no answer to how the system would earn trust or revenue, and nothing that could be run, measured, or falsified.
| absent in 2024 | when and where it arrived |
|---|---|
| Sequence completion as the entire mechanism | continue the table, first exercised in the 2025 first prototype |
| Disjoint pools, independent models, consensus voting | pools and consensus |
| Any concrete, falsifiable experiment | the 2026 synthetic prototype |
| Results-based compensation | paid only on documented early detection |
| The Public Benefit Corporation | incorporated 2026 |
The page also carried scaffolding specific to its audience — a university research partnership, an ethics review board convened with that university, a PR strategy. Those mechanisms did not survive as written; the concerns behind them (ethical oversight, credibility, openness) resurface in the safeguards and publication commitments linked above.
Reading it now
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, everything standardized and eventually published — and left every question of how unanswered. Everything since has been the process of making that page concrete: a mechanism in 2025, a working system and a decisive experiment in 2026, and real claims data next. That progression is traced in the timeline.
It is preserved here as provenance, not as a specification. Where the page and the current system disagree, the page is simply earlier, not wiser.