Skip to content

A Real NOPE Carrier

Status: demonstrated — a real carrier from the prototype database.

Carrier 10705 in pool 3 is one of the 1,500 NOPE carriers in the 2026 synthetic proof of concept: a 45-row ordinary timeline with the PVE TEST dropped in at row 25 and nothing else — no lead-up, no aftermath. NOPE is the specificity control: it teaches the models that without a preceding pattern, the right continuation is silence.

Eight years of ordinary claims

The record reads like anyone's: a thirteen-year-old's routine checkup in 2017, then eight years of colds, urgent-care visits, an eye exam, physical therapy for low back pain, an anxiety diagnosis with counseling and sertraline refills, one emergency visit for abdominal pain. By the final row she is twenty-one and getting a lipid panel. It is an ordinary background history from the synthetic world — because that is all it is, except for one row.

At row 25 — age eighteen, November 2022 — a single claim from an independent laboratory: CPT-82197, the serum veladrin test, the one code every screening read in the prototype greps for. No SIGNAL precedes it; no EARLY or LATE follows it. The excerpt below shows the neighborhood; long descriptions are truncated and rows skip stretches. The leading annotation column (🟡 SIGNAL · 🟠 IMPACT · 🔵 TEST · 🟢 EARLY · 🔴 LATE, blank for base rows) reflects construction marks that are bookkeeping only — no model ever sees them. Here only 🔵 appears; that is the definition of the arc.

PERSONWHENWHEREWHOWHATWHY
13 female2017-04-12POS-11: OfficeNUCC-207Q00000X: Family MedicineCPT-99384: Initial comprehensive preventive medicine evaluation and management…ICD-Z00.129: Encounter for routine child health examination without abnormal findings;
13 female2017-10-15POS-11: OfficeNUCC-207Q00000X: Family MedicineCPT-99213: Office or other outpatient visit for the evaluation and management of an established patient…ICD-J06.9: Acute upper respiratory infection, unspecified; ICD-R05.9: Cough, unspecified;
13 female2017-12-01POS-20: Urgent Care FacilityCPT-99202: Office or other outpatient visit for the evaluation and management of a new patient…ICD-J06.9: Acute upper respiratory infection, unspecified; ICD-R50.9: Fever, unspecified;
18 female2022-04-05POS-11: OfficeNUCC-207Q00000X: Family MedicineCPT-99395: Periodic comprehensive preventive medicine reevaluation and management…ICD-Z00.00: Encounter for general adult medical examination without abnormal findings;
18 female2022-04-05POS-11: OfficeNUCC-291U00000X: Clinical Medical LaboratoryCPT-87389: Infectious agent antigen detection by immunoassay technique… HIV-1 antigen…ICD-Z11.4: Encounter for screening for human immunodeficiency virus [HIV];
18 female2022-04-05POS-11: OfficeNUCC-291U00000X: Clinical Medical LaboratoryCPT-86803: Hepatitis C antibodyICD-Z11.59: Encounter for screening for other viral diseases;
18 female2022-08-12POS-02: Telehealth Provided Other than in Patient's HomeNUCC-207Q00000X: Family MedicineCPT-99213-95: Office or other outpatient visit… synchronous telemedicine service…ICD-R79.89: Other specified abnormal findings of blood chemistry; ICD-Z13.220: Encounter for screening for lipoid disorders;
🔵18 female2022-11-03POS-81: Independent LaboratoryNUCC-291U00000X: Clinical Medical LaboratoryCPT-82197: Serum veladrin / regulatory-mineral ratio, quantitativeICD-E27.A4: Suspected mineralocorticoid-axis dysregulation;
18 female2022-11-14POS-11: OfficeNUCC-2080P0205X: Pediatrics, Pediatric EndocrinologyCPT-99242: Office or other outpatient consultation for a new or established patient…ICD-R79.89: Other specified abnormal findings of blood chemistry;
19 female2023-01-20POS-11: OfficeNUCC-101Y00000X: CounselorCPT-90834: Psychotherapy, 45 minutes with patientICD-F41.1: Generalized anxiety disorder;
19 female2023-01-21POS-11: OfficeNUCC-207Q00000X: Family MedicineATC-N06AB06-207051: Sertraline 50 MG Oral Tablet [#30 DS:30]ICD-F41.1: Generalized anxiety disorder;
20 female2024-03-12POS-23: Emergency Room – HospitalNUCC-207P00000X: Emergency MedicineCPT-99283: Emergency department visit for the evaluation and management of a patient…ICD-R10.9: Unspecified abdominal pain; ICD-R19.7: Diarrhea, unspecified;
21 female2025-05-30POS-11: OfficeNUCC-207Q00000X: Family MedicineCPT-99213: Office or other outpatient visit for the evaluation and management of an established patient…ICD-J06.9: Acute upper respiratory infection, unspecified; ICD-R05.9: Cough, unspecified; ICD-R50.9: Fever, unspecified;
21 female2025-11-10POS-11: OfficeNUCC-291U00000X: Clinical Medical LaboratoryCPT-80061: Lipid panel…ICD-R79.89: Other specified abnormal findings of blood chemistry;

One honest wrinkle: the rows around the TEST can look almost coherent — abnormal blood chemistry flagged a few months before, a pediatric endocrinology consult eleven days after. All of those are base rows, generated before the injection existed; placement is randomized precisely so the TEST cannot correlate with anything in the host record (no placement contamination). The fit is an accident, and a useful one: the control has to hold amid plausible-looking noise, not just sterile quiet.

The dispreferred continuation

NOPE is the mirror image of the GOOD arc. In the training triples, a GOOD prompt ends after the SIGNAL cascade and the preferred continuation fires the TEST; a NOPE prompt is base-only history, the preferred continuation is the ordinary rows that actually follow, and the TEST is the dispreferred path. Same code, opposite lesson: nothing in this record can legitimately predict the test, so a model that emits it here is guessing, not recognizing.

That lesson also gets the most repetitions of any arc — 300 NOPE carriers per pool, against 150 BAD and 50 GOOD. PVE supplies the pattern worth recognizing; NOPE supplies the discipline not to see it everywhere.

What the control bought

In the full results grid, out-of-distribution NOPE prompts — scored only by the four models that never trained on the patient — fired 1.5% of the time, against 100% for GOOD; in-distribution models, which trained on these very timelines, fired 0.7%. Dose-response explains the residue: firing tracks precursor rows present in the prompt, and NOPE prompts contain zero — only a ~1.6% noise floor remains.

Individual models are noisy by design; consensus removes the noise: about 1.4% of single-model runs on NOPE patients fired, falling to 0.13% at at-least-4-of-5 agreement and 0.07% — one patient in 1,500 — at unanimous 5-of-5. Voting across independent pools is the noise filter; this arc is the training that keeps the noise low enough to filter.

Two limits stay attached to those numbers. The data is synthetic and deliberately clean, so a false-fire rate this low is a best case, not a forecast for real claims. And ensemble silence on a patient like this is not a verdict of "no disease" — PRISM reports no negatives; a quiet ensemble simply makes no suggestion.

See also