PRISM

Appearance

Worked Example: FAKE ​

Status: demonstrated — how "earlier" was built, trained, and measured.

FAKE is not a kind of patient — it is training and evaluation material constructed from each BAD carrier by cutting the prompt roughly twenty rows earlier on the same timeline. This article walks that construction end-to-end on a real carrier: how "earlier" was built, trained, and measured.

A construction, not an arc ​

GOOD, BAD, and NOPE are arcs that carriers actually lived through; FAKE names an operation performed downstream on the BAD set. Each of a pool's 150 BAD carriers yields one FAKE example, identical to its BAD example in every respect but one: the prompt — the history the model is shown — ends about twenty rows earlier. The preferred continuation does not change: the same caught-in-time TEST-then-EARLY graft that BAD teaches.

The construction does double duty. In secondary training it is how "surface it sooner" is literally taught: the same lesson, asked from a cut twenty rows earlier. In evaluation it is the only way "earlier" is ever measured: a FAKE prompt ends before the point where the TEST ever occurred, so a model that fires on it has surfaced the screening before the training timeline did.

One generator guarantee makes the cut workable: BAD TESTs are placed at least forty rows into the timeline, so even the deep FAKE cut leaves a real prompt to ask from.

One timeline, two cuts ​

The carrier is the same one the BAD walk-through examines row by row: a 25-year-old male from pool 2, 62 rows, TEST at row 50. In this build, BAD prompts end about 11 rows before the TEST and FAKE prompts about 31 rows before — for this carrier, cuts near rows 39 and 19. One timeline, one question, asked at two depths:

rowswhenmarkwhat happens thereBAD promptFAKE prompt
12025-01-05🟠 IMPACTemergency-room stroke visit✓✓
2–72025-01-05🟡 SIGNAL ×6resistant-BP medications, mineral labs, specialist workups, a lightheadedness visit✓✓
8–19Jan–Marordinary background: refills, office visits, routine labs✓✓
~Mar 2025FAKE cut — the prompt ends here, ~31 rows before the TEST✂
20–39Mar–Augmore ordinary background✓
~Aug 2025BAD cut — the prompt ends here, ~11 rows before the TEST✂
40, 48Aug–Sep🟡 SIGNAL ×2endocrine suppression challenge; serum mineral fractionation
502025-10-06🔵 TESTCPT-82197: serum veladrin / regulatory-mineral ratio
58–59Nov🔴 LATErecurring clearance sessions at a renal-failure facility
60–62Decbackground refills to the end of the record

✓ marks rows included in each prompt; ✂ marks where each prompt ends. The mark column is construction bookkeeping — no model ever sees marks, only the six visible columns of the rows above its cut.

What the earlier cut changes — and what it doesn't ​

In calendar terms, the BAD prompt ends in early August 2025, about two months before the October test; the FAKE prompt ends in early March, roughly seven months before. Everything the ensemble is shown is what a screening system would actually have had in hand that March.

For this carrier the earlier cut costs almost no evidence. Of his nine precursor rows, seven — the stroke and six SIGNALs — landed on the record's opening day, inside both prompts; the remaining two SIGNALs, at rows 40 and 48, sit in the final run-up to the TEST that both cuts remove. That front-loading, with an IMPACT before nearly all of the workup, is not clinical choreography: placement is randomized by design, so a model can key only on the events themselves, never on where they sit.

The read stays binary ​

Each prompt goes to all five models under forced continuation with EOS banned; "fired" means CPT-82197 — the screening test of Primary Veladrin Excess — appears anywhere in the emitted rows. That is the entire read. There is no lead-time score and no row-index credit for firing "more early" (the evaluation guardrails explain why). Earliness lives entirely in the prompt's construction: the record shown ends before the test existed, so if the ensemble surfaces it anyway, the suggestion arrived sooner than the historical record managed.

By that read, across the full grid, FAKE prompts fired 90.0% out-of-distribution — judged by the four models per patient that never trained on it — and 656 of 750 patients drew a unanimous 5-of-5 vote. This carrier's FAKE prompt, seven precursor rows intact, sits comfortably on the firing side of the grid-wide rule — a steered model fires if and only if the prompt contains at least one precursor, and a single row is already a switch.

The honest boundary ​

The ten-point gap from 100% is not failure to recognize evidence — it is prompts with no evidence in them. Because placement is randomized, some BAD carriers' precursors cluster late, and for 63 of the 750 FAKE examples the ~31-row-earlier cut lands before any SIGNAL or IMPACT exists. Those prompts are pure background, indistinguishable from a non-carrier's history, and the ensemble drew unanimous silence on all of them — correct behavior scored against a label that has no signal in it (every miss explained). Conditioned on at least one precursor actually being present, FAKE sensitivity is ~98–100%, in line with GOOD and BAD.

The boundary case is the same result read from the other side: the models fire on the pattern, not the patient. Cut a prompt early enough that the pattern is absent and the correct answer is silence — and silence is what happens.

That is the earliness claim end-to-end. Built: each BAD prompt re-cut about twenty rows earlier, before its TEST ever occurred. Trained: the same caught-in-time continuation, preferred from the shorter record. Measured: a binary fire on those prompts — 90.0% overall, ~98–100% when the cut leaves any evidence, correct silence when it leaves none.

See also ​