Production Scale
Status: vision — the proven 5-pool architecture at population scale.
The 2026 synthetic prototype validated the pool-and-consensus method with five models; the production design scales that same architecture — unchanged in kind — to on the order of one hundred pools. Every number in this article is a design target with a stated rationale, not a description of anything built.
The same architecture, multiplied
The prototype demonstrated the two properties that production scale depends on. First, disjoint pools produce genuinely independent models: no data, weights, or adapters ever cross pools, so agreement between models is independent convergence, not echo. Second, that independence makes agreement a powerful noise filter — in the prototype results, the false-positive rate on control patients collapsed from roughly 1.4% for any single model to 0.07% when all five agreed, while every GOOD carrier drew a unanimous 5-of-5 fire.
Production changes nothing about that unit. One pool, one model, one vote — the design simply asks for more of them. Scaling by adding voters, rather than by making any single model bigger or better, follows directly from the design: the vote is the instrument, individual models are deliberately noisy components, and it is the number of independent perspectives that improves the instrument. Five was the smallest ensemble that could demonstrate cross-pool consensus meaningfully; one hundred is the working target for a real member population. The exact count is not doctrine: it depends on population size, per-pool data volume, and retraining cadence, and it will be settled by the real-data phase, not by this document.
Assignment by stable ID digits
Patients are assigned to pools by trailing digits of a stable patient identifier — for one hundred pools, the last two digits: an ID ending in 47 belongs to pool 47, permanently. In production those identifiers are themselves hash-derived, so pool membership requires no identity knowledge at all (anonymous by architecture).
| property | why it matters |
|---|---|
| deterministic | no sampler, no assignment service, no state to maintain — the ID is the assignment |
| permanent | a patient stays in one pool for life; their growing history never leaks into another pool's training data |
| even | trailing digits of large ID spaces distribute uniformly; each pool holds ~1% of the population with no balancing step |
| coordination-free | one hundred training datasets can be prepared independently, by the same rule, with no communication between them |
The rule's blindness is its virtue. Family members scatter across pools despite shared risks; neighbors scatter despite shared providers; patients with the same condition scatter despite similar histories. Every pool becomes a representative cross-section of the whole population, and no pool can become "the pool that saw the unusual cases." That is exactly the property that made the prototype's cross-pool evaluation honest, arrived at here for free.
What scale buys
| gain | rationale |
|---|---|
| finer consensus gradations | five models yield six agreement levels; a hundred yield a hundred and one |
| rolling daily retraining | retrain one model per day and the whole ensemble refreshes on a ~100-day cycle with no downtime |
| rare-pattern coverage | a genuine pattern recurs across many disjoint pools; a pool-local artifact appears in one |
Consensus resolution. With five models the read is nearly binary — the prototype reported "N of 5" and treated 5-of-5 as its headline. Even at that scale, intermediate thresholds already showed their worth: control-patient false positives fell to 0.13% at ≥4-of-5 agreement before reaching 0.07% at unanimity. With a hundred voters, agreement becomes genuinely graded evidence: thresholds can be set per condition, weighing how invasive the test is against how costly a missed early window is. The read remains what it was in the prototype — a flag, never a probability — but the flag carries far more shading.
Rolling refresh. A hundred-pool ensemble pairs naturally with retraining one model per day: each model refreshes about quarterly, new claims and outcome feedback flow in continuously, and a failed training run degrades one voter out of a hundred instead of the whole system. The staggering also gives the ensemble useful temporal texture — at any moment some voters carry very recent patterns while others hold slightly older ones, so a transient coding fad has to persist across many refresh dates before it can move consensus.
Rare patterns. Each pool of a realistic member population is still large in absolute terms, and a genuine utilization phenotype — however rare — leaves traces in every pool's data. Spurious correlations, by contrast, are pool-local by construction. More independent voters sharpen that distinction. The prototype had to manufacture cross-pool heterogeneity deliberately, generating each pool with a different upstream model; real populations supply that diversity on their own.
What is proven and what is projected
The load-bearing claims — independence from disjoint pools, consensus as a noise filter, silence as the honest default — are demonstrated, not projected. So is the serving economics: one full-context model fits on a single 16 GB consumer GPU (hardware notes), which is what makes one-model-per-card racks a plausible cluster design rather than a hope.
Everything numeric here is projected. One hundred pools, two-digit assignment, the daily cadence — these are targets chosen for clean rationale, and any of them may move when real claims data pushes back. The real-data phase comes first; scale comes after it earns the right to.