PRISM

Appearance

The First Prototype (October 2025) ​

Status: history — superseded by the 2026 synthetic prototype.

In October 2025, two supervised fine-tuning runs on synthetic billing histories produced the first concrete evidence for PRISM's core bet: a language model asked only to continue a patient's claims table learns the general structure of medical care, not the individual patients it saw.

The question ​

Before pools, ensembles, or steering could matter, one question had to survive contact with a GPU: does next-token prediction on formatted billing sequences teach a model anything generalizable? The entire design — sequence completion as the whole task — rests on the answer being yes, and until these runs it was an argument, not an observation.

The setup was deliberately minimal. Roughly 9,790 synthetic patients — about half the dataset then planned — were rendered into an early draft of the timeline format and used for one epoch of full-parameter SFT (what the current design calls primary training) on two open base models, one small and one mid-size. Both were chosen for being openly licensed and, crucially, never instruction-tuned — clean raw material for the exhaustive specialization PRISM asks of a model:

runbase modelhardwaredurationtokens processedfinal train loss
1Qwen3-0.6B-Basefour 16 GB consumer GPUs (homelab)~7.2 h~151 M0.204
2Qwen3-8B-Basetwo rented 96 GB GPUs~7.7 h~151 M0.159

Both runs were uneventful in the best sense — smooth loss descent, stable gradients, no interventions — which itself answered a real question: standard models, standard optimization, no exotic machinery. The larger model's lower loss also showed up as visibly better generations, an early hint that in this domain training loss tracks something real and that added capacity buys added pattern quality.

What the generations showed ​

Validation was a single qualitative probe: one training patient truncated mid-history, with continuations sampled from all four models — trained and untrained, both sizes — under identical seeds.

The untrained models failed in instructive ways. The untrained 0.6B looped, emitting the same procedure code six times in a row. The untrained 8B looked coherent at first — a plausible bundle of acute interventions — until a check of the full record showed it was echoing a nearly identical hospitalization from 2.5 months earlier in the same patient's history. It was pattern-matching the prompt, not modeling care.

The trained models did neither. Both generated novel, medically coherent futures that appeared nowhere in the patient's actual record — a diagnostic workup from the 0.6B, an escalating acute presentation from the 8B — drawn from regularities learned across thousands of patients rather than copied from the context. The 0.6B still over-applied what it learned (a breast-cancer marker ordered for a male patient), a capacity limit noted honestly at the time.

Just as telling was what the models declined to learn. The training data was left deliberately noisy — including anatomically impossible procedures, such as obstetric codes on male patients — and the trained models reproduced none of it. Patterns consistent across the population were learned; errors appearing in one record were smoothed away as noise. That resilience mattered, because real claims data is never clean.

What it did not test ​

This was one bet confirmed, not a system validated. Almost everything that defines PRISM was absent:

absent in October 2025consequence
disjoint pools and an ensembleboth models trained on the same full dataset; independence and consensus were untested
a secondary steering roundnothing nudged either model toward any screening behavior; the second round was then imagined as preference training, an assumption later overturned
an injected conditionthere was nothing to detect, and no firing read to take
an evaluation frameone hand-inspected patient, not a controlled grid with specificity checks and ablations

Superseded in every particular ​

The 2026 synthetic prototype replaced every specific of these runs: a different base model (Qwen3.5-9B-Base, with a reduced vocabulary), five disjoint pools instead of one shared dataset, a per-pool secondary steering round, a fabricated condition to surface, and a full evaluation grid instead of one eyeballed continuation.

What survived is the finding itself. Continued pre-training on billing sequences learns generalizable structure — trained models generate plausible futures instead of echoing recent history, and noise gets smoothed away instead of memorized. Every later result stands on that.

See also ​