PRISM

Appearance

The Daily Model Cycle ​

Status: vision — the production runtime cadence; the 2026 synthetic prototype proved the pieces (independent pools, consensus, gate-on-generation), not this daily cadence at scale.

Production PRISM runs on a single daily rhythm that folds training, screening, and validation into one loop: every day one of the ~100 models is retrained on its own slice of the population, that same slice is screened by the standing ensemble, and the fresh model is mass-validated before it is allowed to replace its predecessor. The fleet reshapes itself between two postures once a day to do it.

One slice, three jobs, one day ​

The population is partitioned into ~100 disjoint pools. Each day the cycle turns to one pool's slice — roughly 1/100 of all members — and that single slice does triple duty:

  • It is the day's training data. The day's replacement model is trained on the fresh slice for its pool, as continuous retraining describes — a rolling refresh that touches every pool once per ~100-day cycle.
  • It is the day's screening batch. The same slice is run through the standing ensemble for suggestions. This preserves the evidence property intact: the slice belongs to one pool, so exactly one standing model has ever trained on those members and the other ~99 judge them out of distribution — the same in-vs-out split the prototype measured, now as the daily default.
  • It becomes the new model's first proving ground. Once trained, the replacement is tested against its own slice plus chunks drawn from other pools.

Training the replacement and screening the slice run concurrently, and they are sized to finish together: the new model comes out of training at about the moment the day's ensemble inference completes.

The daily swap — horizontal, then vertical ​

The same hardware serves two opposite shapes across the day, which is what makes the cadence affordable on one fleet rather than a screening farm plus a separate training farm.

posturewhenshapewhat runs
horizontalthrough the day100 nodes, one model eachthe ensemble screens the day's slice — every member continued by all ~100 models for consensus — while the day's single replacement trains alongside
verticalend of day100 nodes, all hosting the one fresh modelthe new model is pushed its slice and cross-pool chunks at full width, mass-validated in the inference capacity that frees up the instant screening finishes

The day begins spread wide across a hundred different models and ends collapsed onto a hundred copies of one — the fleet pivoting from breadth to depth to spend its end-of-day capacity proving a single model rather than sitting idle.

The daily gate ​

A day closes only on evidence that the new model is at least as good as the one it would replace, judged the way the prototype learned to judge — on free-generation behavior, never training metrics. If the replacement does not clear its predecessor, the predecessor stays and the day's model is set aside. The ensemble therefore never ships a regression: it either improves or holds, one model at a time, every day. Each day ends with three things in hand — one of the hundred slices screened, one newly trained model, and proof it earned its place.

Why this shape, and what it means for detection ​

Three properties fall out of the cadence. Coverage is complete on a rolling basis — every member is screened, and every model refreshed, once per ~100-day turn. Improvement is monotone and cheap to audit — the daily gate is a small, legible checkpoint rather than a periodic mass retrain. And the infrastructure earns its keep in both postures of the day instead of provisioning separately for training and inference.

The cadence also meters detection on its own. Because each member is screened about once every 100 days — roughly 3.65 times a year — suggestions arrive as a steady trickle rather than a flood: on any given day, only the fraction of that day's slice whose pattern has crossed the consensus threshold is surfaced. A condition's undiagnosed backlog is the ceiling of what could ever be found; the daily trickle is the flow that is actually found, set by when each member's signal accrues across their periodic looks — not by any artificial volume cap. The 100-day rhythm is itself the throttle, which is why the system needs no separate alert-fatigue limiter to keep physician-facing volume sane.

Honest status: the prototype ran five models, once, and proved the load-bearing pieces — pool independence, consensus as evidence, and gating on generation. The hundred-model count, the daily swap, and the everyday train-screen-validate loop are production design, not yet built.

See also ​