Continuous Screening
Status: vision — continuous background operation for the real-data phase.
In production, PRISM screens the entire covered population on a rolling background cycle. Nobody orders an evaluation, because the patients who most need screening are often the ones not sitting in front of a doctor.
Nobody triggers it
Preventive care has a structural blind spot: the moment a patient most needs a screening suggestion is frequently a moment when no clinician is looking at them. A patient whose claims record is quietly accumulating a precursor pattern may go years between visits, and by the time symptoms force an encounter, the early-detection window has narrowed or closed. Any screening tool that waits to be asked inherits that blind spot exactly.
PRISM inverts it. Screening runs as batch inference over the whole population, continuously, in the background. The unit of work never changes: assemble a patient's six-column timeline, force a continuation from every model in the ensemble, and count how many independently fire. The 2026 synthetic proof of concept already ran exactly this shape of workload — every patient fired at every model, in batch, across a small fleet. Production turns that experiment into a permanent cycle.
The cycle
Each pass through the population runs five condensed stages:
| stage | what happens |
|---|---|
| Select | choose which patients enter this cycle, ordered by the prioritization below |
| Infer | every ensemble model produces a forced continuation of each selected timeline, in parallel across the GPU fleet |
| Aggregate | count the independent models whose continuation contains the diagnostic TEST code; consensus at or above threshold becomes a candidate suggestion, and silence becomes nothing at all |
| Explain | a separate explanation model drafts the physician-facing narrative from the consensus result alone |
| Deliver | the suggestion and narrative go to the insurer, who routes them to the patient's primary-care physician through existing channels |
The pipeline is deliberately tolerant of failure. Inference nodes drop out and work redistributes; a patient missed in one cycle is simply picked up in the next. No stage records a negative: a patient whose continuation contains no TEST code produces no output, no score, and no entry anywhere that could later be read as "screened and cleared" — silence is not a recommendation.
Who gets evaluated first
Every covered patient cycles through eventually; prioritization only decides the order. A suggestion that lands a week before a scheduled visit can be acted on in that visit; the same suggestion arriving a week after may wait months. Three inputs set the queue:
| input | what it does |
|---|---|
| Temporal alignment | patients with upcoming encounters are evaluated ahead of them, so an actionable suggestion can reach the physician in time for the visit |
| Similarity to confirmed successes | patients whose recent histories most resemble past confirmed early detections are evaluated sooner |
| Baseline randomization | a fixed fraction of every cycle is selected uniformly at random, and periodic sweeps cover the full population regardless of priority |
The similarity mechanism is deliberately mundane. Each patient's recent history — the same six-column text the models read — is reduced to an embedding vector, and nearest-neighbor search ranks the population against the embeddings of patients whose earlier suggestions led to documented early detections. It is a scheduling heuristic and nothing more: similarity never fires a suggestion, never contributes to consensus, and never gates whether a patient is screened — only when. The evidence remains, always, the count of independent models firing.
The randomization floor is the guard against the heuristic's own bias. Any learned prioritization inherits the shape of what it learned from, and patients with fragmented or episodic care — often exactly the patients the system exists for — would otherwise be down-ranked by their own data sparsity. The random fraction and the full-population sweep guarantee every patient is evaluated on a bounded cadence no matter how uninteresting their record looks to the ranker.
No physician initiation — and why that is an equity property
Nothing in this pipeline can be ordered, and that is a design decision with two faces. Operationally, it removes friction: physicians do not need to remember PRISM exists, identify candidates, or submit requests. Their only role is the one that matters — clinical judgment on suggestions received, with full autonomy to act or not.
The second face is equity. A physician-triggered system is a physician-dependent system: coverage would track individual awareness, enthusiasm, panel load, and geography, reproducing the same variation in access the screening gap already reflects. Background operation applies identical analysis to every record, so whether a patient is evaluated depends only on their history, never on who their doctor is or how often they show up. It also dissolves the selective-use question — no one has to answer why some patients were analyzed and others were not, because everyone is.
What is settled and what is not
The mechanics — forced continuation, binary firing, consensus counting, batch execution across a fleet — are demonstrated, not speculative. Everything scheduling-related is not: cycle cadence, consensus thresholds, the weighting among the three prioritization inputs, and the size of the random floor are open operational questions for the real-data phase, to be set with medical oversight and tuned against measured outcomes rather than fixed in advance.
See also
- Cluster architecture — the hardware this cycle runs on
- Continuous retraining — the sibling background loop that keeps the models current
- Production scale — from 5 pools to the vision-scale ensemble
- Clinical decision support — where suggestions land and who decides