The Isolated Signal
Status: demonstrated — enforced in the build, then verified by ablation.
Only one thing in a patient's record is allowed to predict the TEST: the injected phenotype itself. This is the non-negotiable of the 2026 synthetic proof of concept — enforced in three directions during generation, broken once and rebuilt, and verified by ablation at the end.
Why the rule governs everything
The prototype manufactures its own world — 25,000 synthetic patients with a fabricated condition laid over a few thousand of them — and then asks whether the training method learned that condition's precursor pattern. In a manufactured world, that question has a failure mode no amount of good numbers can survive: if the data was built in a way that lets anything else predict the TEST, the models can succeed for the wrong reason — and a contaminated build produces firing rates that look identical to a clean one.
So the generation layer operates under one governing requirement: a carrier must be statistically indistinguishable from a non-carrier except for the injected rows. Anything that violates this — a shared code, a shared trait, a predictable position — is a confound, and a confound anywhere invalidates the result everywhere. Every choice in building the data defaults to whatever makes a positive result harder to explain by anything but the method.
Three directions of contamination
Contamination can enter through what the codes are, who receives them, or where they land.
| direction | the risk | the enforcement |
|---|---|---|
| Code | a phenotype code also appears in ordinary histories, or collides with a real codeset, making a fired TEST ambiguous | every PVE code is fabricated to sit outside all real codesets, and verified absent from every base patient at injection time |
| Selection | carriers share a base-history trait, so models learn the trait instead of the pattern | carriers are chosen uniformly at random, uncorrelated with any feature of the base history |
| Placement | the TEST's position or the lead-up's spacing correlates with something in the host timeline | placement is randomized within a few structural bounds, decorrelated from the host patient |
No code contamination. PVE's codes are deliberately impossible: diagnosis codes with letter-bearing suffixes no real ICD-10-CM code uses, CPT numbers from unused ranges, invented medications. None exists in any real codeset, and the injector verifies at run time that none appears in any base patient. The payoff is that reading a result is a string search, not a judgment call — if CPT-82197 appears in a continuation, it can only have come from the injected pattern.
No selection contamination. Which base patients become carriers is decided uniformly at random — no preference for sicker patients, older patients, or patients with convenient comorbidities. This is the direction the first build got wrong, and it gets the full story below.
No placement contamination. Where the TEST lands, and how the SIGNAL and IMPACT rows scatter through the lead-up, is randomized. This round's only constraints are structural: the TEST never falls in the first twenty rows or the last ten of a timeline, so every training example has a real lead-up and a real continuation, and BAD TESTs sit deep enough that the FAKE cut has room to work. One visible consequence is that carrier timelines can be clinically incoherent — the worked BAD example has an organ-damage event before most of its warning signs. That is by design: the phenotype makes no clinical claim, and choreographing placement to look sensible would itself be a correlation.
The cautionary tale
The first build broke the selection rule on purpose. Nobody knew yet whether the method could work at all, so carriers were preferentially drawn from patients who already had hypertension — an intentional leg-up, giving the injected pattern a plausible-looking host population to sit in. At the time it seemed a reasonable way to de-risk the first attempt.
The consequences were quiet and total. The carrier cohort came out 100% hypertensive, and — because hypertension travels with the whole cardiometabolic cluster — broadly sicker and older than the background population. Models trained on that data appeared to work: they fired on carriers. But they also fired on background patients who merely resembled carriers, because what they had actually learned was "sick hypertensive patient → TEST" — the selection correlate, not the injected pattern. Nothing downstream could disentangle the two explanations, and in an attribution experiment an ambiguous result is a worthless one.
The fix was not a patch. The entire phenotype layer was regenerated from scratch with uniform random selection, and the models retrained on the clean data. The lesson is this article's thesis in one line: a positive result must be attributable to nothing but the method. A shortcut that helps the experiment and a confound that fakes its success are the same object; you only find out which you built after it is too late to matter.
The mirror: strip the rows, get silence
Build-time enforcement is a claim; the design demands verification at evaluation time. If the injected rows really are the only predictive content, removing them from a prompt must silence the models — even a model shown a patient it literally trained on.
The strip ablation ran exactly this test: in-distribution GOOD firing fell from 100% to 0 of 60, and out-of-distribution kinds fell to roughly 1–2%. A model shown a patient it trained on, minus the signal, does not recognize it — it learned the pattern, not the patient. The adapter-off baseline closes the loop from the other side: with the secondary steering disabled, base models fired in only about 0.6% of runs, so the firing itself is attributable to the training, not to base-model fluency.
Together, the build-time enforcement and these ablations are why the prototype's headline numbers mean what they claim to mean. Isolation is also part of what those numbers cost: a signal this clean is an upper bound, not a forecast, and what the prototype does not prove is that real conditions leave patterns this legible.