PRISM

Appearance

Pools and Consensus ​

Status: concept, demonstrated — five pools, five independent models, consensus measured across the full prototype grid.

PRISM never asks what one model thinks. It divides the patient population into disjoint pools, trains one fully independent model per pool, and treats the number of models that independently converge on the same suggestion as the evidence.

One pool, one model, nothing crosses ​

A pool is a permanent, non-overlapping slice of the patient population. Every patient belongs to exactly one pool; every pool trains exactly one model through both training rounds; and nothing ever crosses the boundary. No shared training examples. No shared weights. No shared steering adapters — each pool's adapter is trained on that pool's own model and is never merged or reused elsewhere. The models never communicate, in training or at inference; each continues the same patient timeline alone, unaware the others exist.

This isolation is not a byproduct of the design — it is the design. Agreement between models is evidence only because a spurious pattern in one pool's data can mislead at most one vote. The architecture deliberately gives up the usual benefits of model interaction — shared insights, mutual error correction, merging for efficiency — for a stronger guarantee: when these models agree, that is independent convergence, not groupthink. N models firing is N separate measurements, never one measurement echoed N times.

The 2026 synthetic proof of concept used five pools of 5,000 synthetic patients each, and made the pools deliberately heterogeneous — each generated by a different upstream model and configuration — so the five trained models differ in substance, not just in bookkeeping (the synthetic world).

Consensus as evidence ​

Every model performs the same task on the same patient: continue the timeline under forced generation, and either the diagnostic TEST code appears in the continuation — the model fires — or it does not. One model firing means little on its own; individual models are noisy by design, and PRISM expects them to be wrong in uncorrelated ways. The unit of output is consensus: N of M independent models fired on this patient.

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. Because the models share nothing, the only thing that can make them agree systematically is a pattern that genuinely generalizes across populations.

Two disciplines keep the reading honest. Consensus is a flag, never a probability — "4 of 5 surfaced this" is a count of independent recognitions, not a calibrated likelihood. And silence — no consensus — is never a negative recommendation; a quiet ensemble has said nothing at all (recall, not prediction).

Natural cross-validation ​

Because the pools are disjoint, every patient is out-of-distribution for every model except the one trained on its own pool. Each inference run is therefore also a generalization test: in the prototype, four of the five models had never seen the patient — or any patient from its pool — and had to recognize the pattern purely from what they learned elsewhere.

That is why PRISM never holds out a test set. Every example does double duty: training material for its own pool, evaluation material for the other four. A pattern that exists only in one pool — statistical noise, a data quirk — can earn at most one vote and dies below any consensus bar. Only patterns that manifest across independently constructed populations can produce agreement.

What the prototype measured ​

The prototype ran all 3,250 phenotype examples through all five models — 16,250 completions, zero errors. Counting only out-of-distribution votes — from the four models that never trained on the patient — firing rates tracked in-distribution rates almost exactly, so the pattern was learned, not memorized. Consensus is where the architecture pays off: every GOOD carrier drew a unanimous 5-of-5, while on NOPE — the specificity control, patients whose records contain nothing that can legitimately predict their test — requiring unanimity collapsed the false-fire rate roughly twenty-fold versus any single model's vote. Voting across independent models is the noise filter — a demonstrated result, not a hope. The full firing-rate and consensus tables, every miss explained, and the standing caveat (synthetic, deliberately clean data — an upper bound, not a forecast) live in the prototype results.

From five toward a hundred ​

Five pools is a prototype count: enough to demonstrate the mechanism end-to-end, few enough for one person to build and run. The production vision scales the same architecture to on the order of 100 pools, making consensus finer-grained — "23 of 100" carries more shading than "2 of 5" — without changing anything about how a single vote is produced.

Where the consensus bar sits for a given condition is, in that vision, never a machine-learning decision. Thresholds would be calibrated per condition from the empirical distribution of votes against historical outcomes, set and reviewed under medical oversight, weighing earlier detection against testing burden: PRISM supplies the counts, physicians keep the judgment. The prototype set no thresholds; it reports raw agreement levels only.

See also ​