Sequence Completion
Status: concept, demonstrated — five such specialists were built and evaluated in the synthetic prototype.
A PRISM model does exactly one thing: shown a patient's history as a six-column markdown table, it writes the next plausible rows. There is no chat, no reasoning, no prompt — the whole screening method is assembled from this single, deliberately narrow task.
The whole job
The model receives a billing history rendered as a table — the fixed two-line header, then one row per claim, oldest first: PERSON, WHEN, WHERE, WHO, WHAT, WHY. It produces more of the same table: the rows that would plausibly come next. No system prompt, no instruction, no question is appended to the history. The input is data; the output is more data. The model was trained on raw histories, and at inference it does exactly what it did in training — extend the table, one row at a time.
A screening suggestion fires when the continuation contains the diagnostic TEST code for the condition being screened. Checking is string matching, not interpretation: in the 2026 synthetic proof of concept, the fired-or-silent read was a search for one code, CPT-82197, anywhere in the generated rows. The model is never asked whether the patient should be tested, and it never says so; it writes what comes next, and the presence or absence of a test in that "next" is the entire signal. The models are also never allowed to simply stop — why continuation is forced, and why silence is never read as "no need," is the subject of recall, not prediction.
Why the reframe works
"Should this patient be screened?" is a hard question to hand to a machine. Answering it directly makes the system a decision-maker: it must weigh risk factors, justify a recommendation, defend a threshold, and be audited as though it were exercising clinical judgment.
"What usually comes next in a record like this?" is a different kind of question — the one a language model is natively built to answer. Continuation is what next-token prediction is. If, across a model's training pool, histories resembling this one tended to arrive at a particular diagnostic test, then the statistically plausible continuation of this history contains that test. The screening decision is never made anywhere in the system; it falls out of sequence statistics.
The contrast with the obvious alternative is instructive. A general-purpose model could be prompted — given this patient history, what diagnostic tests would you recommend? — but that framing drags in everything the reframe removes: an instruction to follow, a recommendation to phrase, reasoning to produce, and a conversational model whose every other capability stays live in every response. PRISM's specialists are never asked anything. They are shown a table and they continue it, because continuing it is the only thing they can do.
The reframe also asks nothing of the data that the data does not hold. A claims history contains no lab values, no symptoms as the patient described them, no clinician's impression — only the trail a condition leaves as it gets worked up, its utilization phenotype. Sequence completion operates entirely on that trail. The honest converse applies: if a condition leaves no learnable precursor pattern in claims, this method has nothing to find. The synthetic prototype demonstrated that the machinery works when such a pattern exists; whether real conditions leave comparable trails is the open question the real-data phase tests next.
Intentionally broken models
Catastrophic forgetting — a model losing its general abilities under narrow fine-tuning — is the textbook failure mode of specialization. PRISM makes it the training objective.
The primary training round is continued pre-training: full-parameter training of a general base model on one pool's complete patient histories, and nothing else, until continuing the six-column table is all it can do. The finished specialist cannot hold a conversation, answer a question, follow an instruction, or explain itself. Show it anything other than a patient table and it produces table rows anyway. This is over-specialization pushed well past the point conventional practice calls ruin — deliberately.
The narrowness pays twice. First, capacity: parameters that once encoded dialogue, world knowledge, and instruction-following are repurposed toward the one distribution that matters. Second, surface area: a model that can only emit billing-table rows has no other behavior available to misfire with. There is no latent capability to leak, no persona to manipulate, no instruction-follower to redirect.
A second, far lighter steering round then biases each specialist toward surfacing the TEST when the learned precursor pattern is present in the prompt; how that round works — including the preference-training approach that failed first — is that article's story. The division of labor is the point here: the primary round builds a model that can only continue the table, and the secondary round shapes which continuation it prefers.
What simplicity removes
Most of the failure modes that dominate discussion of deployed language models require machinery PRISM never builds.
| common failure mode | why it has no purchase here |
|---|---|
| prompt injection | there is no prompt to inject into — the input is rows of standardized codes, and any text inside a row is just more sequence to continue |
| jailbreaking | there is no refusal boundary and nothing behind one; the model has no capability beyond emitting table rows |
| unfaithful reasoning | there is no reasoning chain — nothing is generated between the history and the continuation |
| hallucinated explanations | the model never explains anything; checking its output means checking whether a code appeared, not auditing an argument |
| prompt-engineering drift | there is nothing to engineer — the "prompt" is the patient's history, assembled the same way every time |
None of this claims the system is failure-free. It has failure modes of its own — individual specialists are noisy by design, and the prototype results account for every miss and false fire mechanically. The claim is narrower: the failures that remain are simple and inspectable. A specialist either continues a history with the TEST code or it does not, and because no single continuation is ever the verdict, the unit of evidence is consensus across independent pools.
See also
- Two training rounds — how the specialists are actually made, and the training detour that taught the most
- Recall, not prediction — forced continuation, and why silence is never a negative recommendation
- Pools and consensus — why one model's continuation is never the answer
- Patient timeline format — the six columns a model sees, and everything it never does