Glossary · Reference for every page

A small vocabulary, chosen on purpose.

concept

Every load-bearing PRISM term, defined plainly enough to stand alone. Most of the vocabulary falls into three families: format terms describe what a model actually sees, construction terms describe how the training material was made, and method terms describe how the system reasons and reports.

One rule spans every family

Construction vocabulary is bookkeeping. Marks, arcs, pool assignments, patient identifiers — none of it is ever visible to a model. If a term describes how data was made rather than what a model sees, no model has ever seen it.

First, the name

The system itself.

concept

PRISM: Predictive Recommendations for Improved Screening in Medicine: a system that turns the anonymized insurance-claims histories an insurer already holds into early-screening suggestions — flags to a physician that a specific, established diagnostic test may be worth ordering now. → Overview
recall instrument: What PRISM is instead of a predictor: it surfaces candidates for a test, and its silence carries no information. Despite the word "Predictive" in the acronym, nothing in the system predicts, diagnoses, or scores. → Safeguards
constructive-only: The architectural guarantee that PRISM can only suggest a test, never deny one; no pathway exists to emit "don't test". → Safeguards

Format terms

What a model actually sees.

concept demonstrated

PERSON · WHEN · WHERE · WHO · WHAT · WHY: The six columns of a patient timeline and the only fields any PRISM model ever sees, in training or inference: age and sex, date, place of service, provider taxonomy, procedure or product, and diagnosis — one row per medical event. No name, address, or free text exists anywhere in the pipeline. → How it works
header: The fixed two-line markdown table header that begins every example; served from the database and cached, never hard-coded. → How it works

Construction terms

How the prototype's training material was made.

concept demonstrated

synthetic prototype: The 2026 synthetic proof of concept: five pools of 5,000 synthetic patients, Primary Veladrin Excess (PVE) injected, five models trained and evaluated end-to-end. The evidence backbone of everything this site reports. → Evidence
PVE: Primary Veladrin Excess, the fabricated condition at the center of the synthetic prototype. Deliberately fictional, with no clinical mechanism, so that results are attributable only to the injected pattern. → Evidence
carrier: A base patient who received injected PVE rows — roughly 2,500 of the prototype's 25,000 synthetic patients; the rest are clean background. → Evidence
arc: A carrier's injected storyline — GOOD, BAD, or NOPE — each built to teach the ensemble a different lesson. → Evidence
GOOD: The "caught in time" arc: SIGNAL lead-up, then TEST, then EARLY, with no IMPACT damage. → Evidence
BAD: The "missed until it hurt" arc: precursor SIGNALs and accruing IMPACT damage, a TEST that arrives late, then ongoing LATE treatment. → Evidence
NOPE: The specificity-control arc: a bare TEST with nothing in the record that could legitimately predict it; teaches models when not to fire. → Evidence
marks: The five construction labels — SIGNAL, IMPACT, TEST, EARLY, LATE — that slice injected rows into training material. Bookkeeping only; no model ever sees them. → Evidence
SIGNAL: The mark for the diverse precursor lead-up — workups, labs, scattered specialist visits, vague symptom encounters — that precedes the TEST in GOOD and BAD arcs. → Evidence
IMPACT: The mark for accrued end-organ damage, found only in BAD arcs — strokes, infarcts, crisis admissions — the cost of a missed diagnosis. → Evidence
TEST: The diagnostic screening event and its mark; in the prototype, the fabricated lab code CPT-82197, which exists in no real codeset. Its appearance in a continuation is what "fires" means. → Evidence
EARLY: The "caught in time" outcome mark: after a timely TEST, the storyline settles into recurring refills of a simple daily tablet. → Evidence
LATE: The "missed until it hurt" outcome mark: recurring institutional clearance sessions, the ongoing burden a timely TEST would have avoided. → Evidence
FAKE: Constructed training and evaluation material, not a kind of patient: a BAD example cut about 20 rows earlier, so firing on it means surfacing the TEST before the data did. It exists so that "earlier" has an operational, binary meaning. → Evidence

Honest limits

Every count in this family — five pools of 5,000, roughly 2,500 carriers, a condition and a code that exist nowhere — describes fabricated data, on purpose. That isolation is what made the 2026 results attributable only to the injected pattern. The rates those results report are an upper bound measured on deliberately clean synthetic data, not a forecast of real-world performance. Read the results and their limits →

Method terms

How the system reasons and reports.

concept demonstrated

pool: One of the disjoint training populations. One pool trains one model; no data, weights, or adapters ever cross pools. → How it works
ensemble: The set of independently trained models — five in the prototype — that each continue the same patient and vote. → How it works
primary training: Full-parameter continued pre-training of a base model on one pool's complete histories, until continuing the six-column table is all it can do. → Researchers
secondary training: The per-pool steering round: a LoRA adapter trained by pure SFT (supervised fine-tuning) on chosen continuations, applied to that pool's own primary model and never shared across pools. → Researchers
utilization phenotype: The pattern a developing condition leaves in billing data — the thing PRISM actually learns to recognize. PVE is one manufactured instance of one. → How it works
fires / firing: The TEST code appearing anywhere in a model's forced continuation. The read is binary: present or absent. → How it works
consensus: N of M independent models firing on the same patient — the unit of evidence. Reported as a flag ("4 of 5 surfaced this"), never a probability. → How it works
OOD (out-of-distribution): Every example trained exactly one pool's model, so it is unseen data for the other four; only those four count as evidence. → Evidence
silence: The absence of consensus on a patient. Explicitly not a negative recommendation; never reported as "no need". → Safeguards

Easy to conflate

Four distinctions worth keeping.

concept

Marks vs. arcs

They operate at different altitudes. SIGNAL, IMPACT, TEST, EARLY, and LATE label individual rows; GOOD, BAD, and NOPE describe the whole storyline those rows compose. Both exist only in the construction pipeline — no model ever sees either, and nothing in a rendered timeline betrays them.

FAKE is not a fourth arc

There are no FAKE patients. A FAKE is an example manufactured from a real BAD carrier by cutting the prompt about 20 rows earlier, and it exists so that "earlier" has an operational, binary meaning: did the ensemble surface the TEST before the data did.

TEST is three things at once

A mark on a row, an event in a storyline, and a concrete billing code — but at read time it collapses to a single question: does the code appear in the continuation. Everything downstream of inference treats it as that one binary.

Silence is not NOPE

NOPE is a training arc — records that contain nothing predictive, teaching models to keep continuing normally. Silence is a runtime outcome — the ensemble not reaching consensus on some patient. The first is a lesson; the second is an absence, and it is never converted into a recommendation.

Words avoided on purpose

The vocabulary keeps wrong intuitions out.

concept

Part of the vocabulary exists to keep wrong intuitions out. Each of these words was chosen over a more familiar one because the familiar word imports an assumption PRISM's design explicitly rejects.

We say	Never	Because
fires	~~predicts positive~~ · ~~classifies~~	Prediction language smuggles in the idea that the output is a probability with a complement — that a non-fire means something. It does not: PRISM is a recall instrument, and forced continuation exists precisely so that no model can be read as declining.
consensus	~~confidence~~ · ~~score~~	The count of independently trained models that fired is a flag, not a calibrated probability.
pool	~~fold~~ · ~~split~~ · ~~shard~~	Those words import cross-validation intuitions — hold-out sets, per-model precision as a verdict — that the evaluation design explicitly rejects.
carrier	~~case~~ · ~~positive patient~~	A carrier is defined by construction — rows were injected into that record — not by any clinical judgment. In a synthetic world there is nothing to diagnose, only a pattern to place and later recognize.

One last convention

How every claim wears its status.

concept

PRISM's documentation — and this site — mixes locked experimental results with production intent, and the two must never blur. Four tags say how much weight to put on any claim.

concept: A durable design principle — how PRISM is meant to work, whatever the scale or phase.
demonstrated: Validated end-to-end in the 2026 synthetic prototype, on deliberately clean fabricated data. → Evidence
vision: Production design intent, not yet built — never written as though it already exists. → Vision
history: A record of past work, superseded, kept for provenance. → Story

This glossary itself is concept — these words mean the same thing everywhere on this site, whatever the status of the page using them.