The Synthetic Prototype

Status: demonstrated — completed mid-2026; results in prototype results.

The 2026 synthetic proof of concept — "the synthetic prototype" throughout this documentation — is the completed experiment at the center of PRISM: a controlled, deliberately artificial world built to answer one question about the method, cleanly enough that a positive answer could mean only one thing.

A deliberately artificial experiment

The prototype does not attempt to mimic clinical reality, and it does not claim to prove real-world efficacy. It manufactures one condition that does not exist — Primary Veladrin Excess (PVE), a plausible-sounding disorder with deliberately no clinical mechanism — and injects its billing footprint into a synthetic world of 25,000 generated patients across five disjoint pools. Every PVE code is fabricated to sit outside all real codesets, so when a model's continuation contains the screening TEST code, CPT-82197, the read is a string match, not a judgment call.

The artificiality is the point. Real claims data would leave a positive result open to endless alternative explanations — the models keyed on age, on comorbidity, on one insurer's coding habits. The prototype's construction discipline (see the isolated signal) exists so that the only thing in any patient's record that can predict the TEST is the injected pattern itself. A positive result can then only mean the method worked.

The one question

Can the training method steer the ensemble to surface the screening TEST earlier in a patient's timeline than it appeared in the training data, learned purely from the utilization pattern that precedes it?

Every clause is load-bearing. "The ensemble" means five independently trained models, one per pool, whose agreement — never any single output — is the signal (pools and consensus). "Surface the TEST" means the code appears in a forced continuation: the model fires, it does not "predict positive" (recall, not prediction). "Earlier than it appeared in the training data" is operationalized by exactly one device — FAKE prompts, the BAD training lesson asked again from a cut roughly twenty rows earlier, so the prompt ends before the point where the TEST ever occurred; a model that fires on one has surfaced the screening before the data did (the FAKE cut). And "learned purely from the utilization pattern" is the isolation guarantee above, enforced in three directions and mirrored by a strip test.

The wager outside the experiment

One belief lives outside the experiment, deliberately untested by it: that some real conditions leave precursor patterns in claims data conceptually like PVE's — utilization phenotypes that models could learn from billing histories without anyone ever naming or describing them.

The prototype neither tests this wager nor could. Its job was narrower and prior: build one clean, manufactured instance of such a pattern and establish whether the machinery — disjoint pools, two training rounds, forced continuation, cross-pool consensus — can learn it and act on it at all. If the method failed even here, on a signal isolated by construction, there would be nothing worth taking to real data. Because it succeeded, the wager is worth placing — on real claims data, in the next phase.

The verdict, in one breath

On the controlled synthetic world, the method works — decisively. Five independently trained models, each shown patients it never trained on, surface the screening TEST when — and only when — the injected precursor pattern is present; they surface it earlier than the training timeline did; and they agree with each other when they do. Strip the injected rows back out and they fall silent — the exact win-condition the experiment named in advance. And every residual miss and false fire is explained mechanically rather than hand-waved: across the entire dataset, a steered model fires if and only if the prompt contains at least one precursor row.

The numbers behind that sentence — the full grid of 3,250 patients × 5 models = 16,250 runs, firing rates by kind, the consensus tables, the steering baseline, the strip ablation, and the dose-response curve — live in prototype results.

What it does not prove

• The data is synthetic and deliberately clean. The signal is isolated by construction — that is the whole point, since it makes a positive result attributable — but it also means the measured rates are an upper bound, not a forecast of real-world performance.
• It proves the method works given an analogous precursor pattern. It does not establish that real conditions have such patterns, or how strong or learnable they are. That is the explicit leap of faith above, to be tested next on real claims data with retrospective evaluation.
• The smaller sub-analyses (dose-response, strip ablation) are directional; the full-population grid is the robust result.
• Single-model behavior is noisy by design; only the cross-pool consensus counts as evidence.

No number from the prototype should be quoted without this section attached to it.

Map of the prototype section

article	what it covers
The synthetic world	5 pools × 5,000 patients; deliberate cross-pool heterogeneity; carriers and clean background
Primary Veladrin Excess	the fabricated condition in full: fiction by design, the arcs, the actual code inventory
The isolated signal	the one non-negotiable: three contamination directions, a cautionary tale, the strip mirror
Building the data	the generation pipeline and how training material is sliced from it
Training runs	the base model, both training rounds as actually run, the artifacts per pool
Running the ensemble	the five-node fleet and how the 16,250-run grid was executed and captured
Results	every findings table, every miss explained, and the limits restated
A GOOD carrier	a real caught-in-time timeline, walked through row by row
A BAD carrier	a real missed-until-it-hurt timeline
A NOPE carrier	the specificity control: a bare TEST that nothing in the record should predict
The FAKE cut	how "earlier" is constructed, and what firing on it means
Hardware notes	measured throughput per GPU class; why one 16 GB consumer card is a full node