Primary Veladrin Excess
Status: demonstrated — the condition is fictional; its results are real.
Primary Veladrin Excess (PVE) is the fabricated condition at the center of the 2026 synthetic proof of concept: a disease that does not exist, invented so that detecting it can mean only one thing.
Fiction by design
PVE sounds like a real condition — a mineralocorticoid-flavored disorder of an invented regulatory mineral, "veladrin" — and that is as far as the fiction goes. It has no clinical mechanism, deliberately. It encodes no theory about endocrinology, no hypothesis about what real precursor cascades look like, no claim that any actual disease behaves this way. It is one clean, manufactured instance of a learnable precursor pattern: a spread of ordinary-looking encounters that reliably precedes one specific diagnostic test.
That emptiness is the point. The prototype asks whether the training method can find and act on a precursor pattern at all, and answering that requires a pattern whose presence, shape, and placement are fully known — something no real condition can offer. Because PVE carries no clinical content, a positive result cannot be credited to medical plausibility; because the signal is strictly isolated from everything else in the data, it cannot be credited to an accident of construction either. What the fiction does not establish is stated just as plainly in the results: whether real conditions leave analogous patterns in claims data is the explicit leap of faith, tested next on real data.
PVE exists only as rows. It is a pure utilization phenotype — a pattern of billing events laid over a randomly chosen subset of the ordinary patients (the carriers) in the synthetic world. Every injected row carries one of five construction marks — SIGNAL, IMPACT, TEST, EARLY, LATE — which downstream tooling uses to slice timelines into training material. The marks are bookkeeping only. No model ever sees them, in training or at inference; a model sees exactly the six ordinary columns of the timeline format, where an injected row is indistinguishable in kind from any other claim.
Codes that exist nowhere else
Every PVE code is fabricated to sit outside all real codesets: diagnosis codes with letter-bearing suffixes no real ICD-10-CM code uses (E27.A5, R42.X3), CPT numbers drawn from unused ranges (82197, 90981), and an invented medication with a made-up ATC class and RxCUI. None of them appears in any base patient, and none collides with the real code systems the background data is built from. The consequence is operational: whether a model surfaced the TEST is a string match — a grep, not a judgment call — and the convention enforces the no-code-contamination direction of the isolated-signal requirement.
The pivot: one TEST, two futures
| mark | WHAT | WHY | setting |
|---|---|---|---|
| TEST | CPT-82197: Serum veladrin / regulatory-mineral ratio, quantitative | ICD-E27.A4: Suspected mineralocorticoid-axis dysregulation | independent laboratory, exactly once per carrier |
| EARLY | ATC-C03DZ01-211582: veladrostat 12.5 MG Oral Tablet [#90 DS:90] | ICD-E27.A5: Primary veladrin excess | pharmacy, recurring refills |
| LATE | CPT-90981: Extracorporeal regulatory-mineral clearance session, per treatment | ICD-E27.A5: Primary veladrin excess | End-Stage Renal Disease Treatment Facility (nephrology), recurring |
The TEST appears exactly once in every carrier's timeline — all ≈2,500 of them across the five pools — and everything forks on when it arrives. Caught in time, the future is EARLY: a well-managed patient who just keeps refilling a simple pill, ninety days at a time. Caught late, the future is LATE: a recurring, dialysis-flavored facility burden — about 39 clearance sessions per missed carrier in this build, 29,246 LATE rows across the 750 BAD carriers. The contrast is deliberate and legible even to a system that understands nothing of medicine: caught early means a pill; caught late means an ongoing institutional burden. That asymmetry, written directly into the billing record, is what makes the early TEST worth learning to surface.
SIGNAL — the lead-up
The precursor cascade is deliberately diverse — roughly 22 distinct record types across four loose families — so that no single event is the pattern:
| family | records |
|---|---|
| resistant blood-pressure workups | CPT-93785 ambulatory blood-pressure monitoring; refills of the fabricated antihypertensives hydralozide and clazapamide |
| mineral / electrolyte labs | CPT-80261 broad metabolic survey; CPT-82196 serum regulatory-mineral fractionation; CPT-83541 24-hour urine mineral panel |
| scattered specialist evaluations | cardiology rhythm strips; GI motility; neurology conduction; dermatology deposition; rheumatology myopathy; urology and gynecology mineral-axis assessments; chiropractic |
| vague symptom visits | ICD-R53.A0 persistent unexplained fatigue; ICD-R42.X3 recurrent positional lightheadedness |
The variety paid off measurably: the trained ensemble fires on the presence of the pattern, not its volume — a single recognizable precursor row in the prompt was enough to fire over 90% of the time (results).
IMPACT — the accrued damage
IMPACT rows appear only in BAD carriers: the harm that piles up while the condition goes uncaught. Seven event types:
| event | code |
|---|---|
| accelerated chronic kidney disease, office follow-up | ICD-N18.A4 |
| prolonged renal support | ICD-PCS-5A1D8ZZ, with ICD-N18.A4 |
| stroke, emergency visit | ICD-I63.A2 |
| myocardial infarction, emergency visit | ICD-I21.A2 |
| metabolic-crisis admission | ICD-E87.A9 |
| heart-failure admission | ICD-I50.A1 |
| end-organ retinopathy | ICD-H35.A1 |
The arcs
Each carrier receives one of three storylines — an arc — woven into its ordinary history. A fourth kind of training example, FAKE, is constructed later and has no patient of its own.
| kind | shape | what it teaches | walk-through |
|---|---|---|---|
| GOOD arc | SIGNAL lead-up → TEST → EARLY; no IMPACT | when the pattern appears, fire the TEST — then the pill | a real GOOD carrier |
| BAD arc | SIGNAL plus a few IMPACT → late TEST → LATE | prefer an early TEST→EARLY graft over the real drift into LATE | a real BAD carrier |
| NOPE arc | a bare TEST and nothing else | stay silent — nothing in the record can legitimately predict this TEST | a real NOPE carrier |
| FAKE (construction) | a BAD example asked again from a cut ~20 rows earlier | the same preference from less history — "surface it sooner," literally trained | the FAKE construction |
NOPE is the specificity control: same TEST code, zero lead-up, so a model that fires on it is guessing, not recognizing. FAKE is worth restating because it is easy to misread as a patient type — it is not. It is training material derived from BAD carriers, and it is also how "earlier" is operationalized end to end: a model that fires on a prompt cut before the TEST ever occurred has surfaced the test before the data did. How the arcs become prompt/chosen/rejected training triples belongs to three-pattern learning; the per-pool carrier counts belong to the synthetic world.
Placement follows a few hard constraints — the TEST never lands in the first 20 rows or the last 10 of a timeline, and BAD TESTs sit at least 40 rows deep so the FAKE cut has room to work — and is otherwise randomized. No arc is clinically choreographed: a BAD carrier can log an IMPACT before most of its SIGNALs, and that is by design. The phenotype makes no clinical claim, and tying placement to anything in the host patient would be exactly the contamination the prototype exists to rule out.
See also
- The isolated signal — the non-negotiable that makes PVE results attributable
- The synthetic world — the pools, populations, and carrier counts PVE is injected into
- Three-pattern learning — how the arcs become training material
- Prototype results — what the ensemble actually did with all of this