The Synthetic World
Status: demonstrated — this is the world the prototype was trained and judged on.
The 2026 synthetic proof of concept was trained and judged entirely inside a manufactured world: 25,000 synthetic patients across five deliberately different pools, with one fabricated condition injected over a small, randomly chosen subset. The world has exactly two layers — ordinary background and injected signal — and the boundary between them is what the whole experiment rests on.
Two layers
The first layer is the base histories: five pools of 5,000 synthetic patients each. Every patient is an ordinary medical timeline in the six-column claims format — chronic-condition management, acute events, routine screenings, prescription refills. What the base layer contains none of is the condition under study: zero phenotype rows, and the phenotype's codes appear in no base patient and no real codeset.
The second layer is the injected phenotype: Primary Veladrin Excess, a fabricated condition laid over a randomly selected subset of base patients as additional rows woven into their timelines. This layer is where the signal the experiment cares about is created, and where nearly all of the design attention went. The rule governing the boundary between the layers is the prototype's one non-negotiable: the only thing in any patient's record that may predict the screening TEST is the injected pattern itself.
Construction marks tag the injected rows for bookkeeping — which rows are lead-up, which is the TEST, which are outcomes — but no model ever sees them. They exist so the data pipeline can slice the world into training material.
Five pools, deliberately different
The pools are not five samples from one generator. Each pool's base histories were produced by a different upstream model and configuration, on purpose. Within a pool the data-language is internally consistent; across pools it diverges by design — different row densities, description phrasings, coding habits.
That heterogeneity does a specific job. Real claims data varies by population, insurer, and coding culture, and the five pools stand in for that variety. Because each ensemble model trains on exactly one pool and shares nothing with the others — no data, no weights, no adapters — the heterogeneity is what makes their agreement mean something. Five models trained on five draws from the same generator would be five copies of roughly one opinion; five models trained on five different data-languages that still converge on the same patient are much closer to independent witnesses.
Who carries the condition
Carriers were chosen uniformly at random from each pool's 5,000 patients — no clinical criteria, no correlation with anything in their base history. A carrier is statistically indistinguishable from a non-carrier except for the injected rows. That blandness is load-bearing: the first build selected carriers hypertension-first, the models learned the wrong pattern, and the dataset had to be rebuilt from scratch — the cautionary tale is told in the isolated signal.
Each carrier receives one arc:
| kind | per pool | what it is |
|---|---|---|
| GOOD | 50 carriers | signal lead-up → TEST caught in time → simple ongoing treatment |
| BAD | 150 carriers | signal plus accrued damage → late TEST → costly recurring facility care |
| NOPE | 300 carriers | a bare TEST with no lead-up at all — the specificity control |
| FAKE | 150 examples | not patients: training material constructed downstream from the BAD set |
FAKE earns the clarification in its last cell: it is not a fourth kind of patient. FAKE examples are built by the data pipeline from BAD carriers, cutting the prompt roughly twenty rows earlier so the same lesson is asked from less history. That construction is how "surface the test sooner" is both trained and, later, measured.
The totals: 500 carriers per pool of 5,000 — roughly 2,500 carriers and 3,250 phenotype examples across the whole world. The other ninety percent of patients are clean background, and they are not filler: they are what teaches each model what ordinary looks like, and they are the population on which silence can be verified.
What this world is not
The world makes no claim to clinical realism. The phenotype encodes no medical mechanism, the carrier fraction was chosen for experimental convenience rather than epidemiological accuracy, and the base histories carry known generation quirks — hallucinated medication codes, thinner-than-specified timelines — that were accepted rather than fixed (building the data). None of that undermines the experiment, because the experiment never asked the world to be realistic; it asked it to be clean, so that a positive result could only be attributed to the method. The matching honest limit runs the other way: results earned in a deliberately clean world are an upper bound, not a forecast.
See also
- Primary Veladrin Excess — the fabricated condition in full
- The isolated signal — the non-negotiable, and the hypertension-first cautionary tale
- Building the data — the generation pipeline and its accepted quirks
- Pools and consensus — why independence is the architecture