Open Collaboration
Status: vision — the long-term network; nothing here exists yet.
The long-term ambition is a network in which many organizations improve a shared screening ensemble by exchanging trained models — and never, under any circumstances, patient data. Of everything in the PRISM vision this is the furthest from the working prototype, and this article should be read with that distance in mind.
Share models, never data
An organization that trains a PRISM model on its own claims produces a file of numerical weights. Those weights encode learned regularities — that certain sequences of billing codes tend to precede a diagnostic test — but they are an abstraction of the training data, not a copy of it. There is no straightforward path from weights back to an individual's record, and PRISM does not need to rest the whole argument on that irreversibility alone: the training data is anonymous by architecture before any model sees it. A model trained on nameless, locationless, identifier-free timelines has nothing identifying to leak even in the worst theoretical case. The two protections stack, and that stacking is what makes sharing weights defensible where sharing data never would be.
This separation is the enabling move. Organizations that could never pool patient records — for privacy, regulatory, or competitive reasons — can exchange what their data taught, because what their data taught is all a model contains.
Merging without crossing pools
Independence between pools is what makes consensus mean anything: agreement is only evidence when the agreeing models could not have copied from each other (see pools and consensus). A collaboration that casually merged everything into one super-model would destroy the very property being scaled.
So merging is strictly pool-wise. Every participating organization partitions its population into the same pool structure, and a pool-N model from one organization merges only with pool-N models from others. The invariant survives: no patient's data ever influences more than one pool's model, anywhere in the network. A merged pool-N model has still never seen a patient from any other pool — when it votes on one, that is still genuine out-of-distribution recognition, now informed by several populations' pool-N patients instead of one. The ensemble gets richer without getting less independent.
An honest note: merging independently trained model weights is an active research area, not a solved procedure. Which techniques preserve what this architecture needs is an open question that only the production phase can answer.
Open code, licensed weights
The intended licensing structure separates three layers with different terms.
| layer | terms | rationale |
|---|---|---|
| the codebase — data pipeline, training, ensemble, consensus, evaluation | open: inspect, modify, self-host | there are no secret algorithms; verifiability is the point, and anyone may check exactly what the system does |
| trained model weights — the collective ensemble | licensed: access requires active participation | the value of the network lives in the weights; access to everyone's learned patterns is earned by contributing your own |
| patient data | never shared, never requested, never leaves the organization | the foundation everything above depends on |
Participation means continuous contribution, in two currencies. First, models: each participant keeps retraining on its own population and feeds the fresh weights back on the network's cadence, so the collective ensemble is a living thing rather than a snapshot. Second, improvements: technical advances made during implementation — faster data preparation, better training procedures, sturdier deployment — flow back to all participants rather than accumulating as private edges.
An organization could take the open code and run entirely alone on its own population. The license simply prices the alternative honestly: access to patterns learned from populations you will never insure costs contributing the patterns learned from the one you do.
Network effects, compounding early
Each new participant brings populations, provider networks, and coding habits the rest of the network has never seen — exactly the heterogeneity that makes independent agreement meaningful. The 2026 synthetic proof of concept had to manufacture that diversity deliberately, generating each of its five pools with a different upstream configuration; a real network would get it for free, and in greater variety than any single carrier's membership can supply. More genuinely independent perspectives make consensus stronger evidence and filter noise harder — the prototype's own results showed false positives collapsing as the required agreement rose.
The benefits compound for those who arrive early. An early participant contributes weights trained on its own members and then gains from every later arrival — new pattern diversity, stronger consensus — without retraining or additional effort. Early participants also shape the standards, cadence, and governance that later ones inherit. None of this is a moat against joining late; it is a reason not to.
Compete on implementation, collaborate on pattern recognition
Insurers reflexively treat capabilities as proprietary. Pattern recognition is a poor place for that instinct: no single carrier's population is diverse enough to learn every presentation of a precursor pattern, and a hundred organizations independently rebuilding the same recognition machinery is pure waste. The reframe this network proposes is to collaborate on recognition — the shared, merged ensemble — and compete on implementation: how well suggestions reach physicians, how reliably patients follow through, how efficiently the screening-to-diagnosis pipeline runs. Recognition improves for everyone as the network grows; differentiation moves to where organizations actually differ.
The honest distance
Nothing in this article exists. There is no repository, no merging pipeline, no license text, and no second organization. Before any of it matters, PRISM must first work on real claims data, then scale past the prototype's five pools toward the production architecture with its continuous retraining rhythm, and then persuade at least two organizations that the argument above holds. This is the last part of the vision to be built because it is the part that must be earned.
See also
- Pools and consensus — the independence this network is designed to preserve
- Anonymous by architecture — why sharing weights is safe in the first place
- Production scale — the ensemble this network would grow
- Continuous retraining — the cadence contributions would follow