Vocabulary¶
Short definitions of the terms used in this report. The TypeSafe terms follow the TypeSafe docs.
TypeSafe and Jev¶
- System One / System Two
- After Kahneman's Thinking, Fast and Slow. A System One model makes fast, focused judgments and returns typed answers, not text. System Two here means a slower, deliberate reasoning step: a large language model (Claude Sonnet 5) or a human, called only when System One is unsure.
- Jev
- TypeSafe's flagship System One model. This report pins
jev-1.13.0. It is priced per input token; output tokens are free. - State
- The content Jev evaluates: a string, JSON object or array. Here it is a small JSON object that code builds from FHIR data and free text. Questions refer to parts of it by path, for example
`nursing_note`. - Question / primitive
- One typed judgment about the state. All questions in a request see the same state and are answered independently and in parallel. There are three primitives:
| Primitive | Asks | Returns | Example here |
|---|---|---|---|
| Choice | Which one of these options? | choice, a probability per option, confidence |
Who should handle this patient message? |
| Score | Where on this ordered scale? | score (a probability-weighted level, can be fractional), probabilities per level, confidence |
How worried should the ward team be (levels 0–3)? |
| Noul | Is this statement true? | noul: the probability of yes (0–1) |
Does the note describe new confusion? |
- Criteria
- The answer space: Choice options, Score levels, or optional true/false definitions for a Noul. Clear criteria matter more than clever instructions.
- Confidence
- A 0–1 summary of how concentrated a Choice or Score probability distribution is. It describes the model's certainty, not correctness. Nouls have no separate confidence: a value near 0.5 is itself the uncertainty signal.
- Fan-out
- Asking many questions over one state in a single request, for example one Choice per medication. Latency barely changes as questions are added.
- Confidence gate / escalation
- Rules in code that act on an answer only when it is confident, and otherwise send it to System Two. Examples: route confidence < 0.6, or a Noul between 0.2 and 0.8.
- Jaggedness
- TypeSafe's term for Jev's known weak spots: literal reading, arithmetic, dates, multi-hop reasoning ("indirection"), large irrelevant state, and generation.
Task categories¶
Decision shapes from the TypeSafe use-case map, as used in the scenarios:
| Category | Meaning | Where it appears |
|---|---|---|
| Classification | One known category should win | Type of deterioration (ward) |
| Detection | Probability that one property is present | New confusion, infection, red flag, safeguarding |
| Scoring | Place something on an ordered rubric | Nurse concern, message urgency, quality of explanations |
| Routing | A category selects the next code path | Inbox queue |
| Ranking | Order items by a semantic signal | Ward huddle priority list |
| Verification | Check an artefact for specific failure modes | Allergy conflict, duplicate therapy, interaction |
| Structured data extraction | Recover known fields from unstructured text | Status of each medication at discharge |
Evaluation terms¶
- Reference label
- The expected answer for each case, written by Claude Opus 5 and committed before Jev ran. This is not a clinician gold standard.
- Ambiguous
- A reference label that could reasonably go either way. Accuracy is reported with and without these.
- Brier score
- Mean squared error between a probability and the 0/1 truth; lower is better, and 0 is perfect. It measures calibration as well as correctness.
- MAE
- Mean absolute error between a fractional Score and the reference level.
- Under-triage / over-triage
- Assigning a less urgent (under) or more urgent (over) band than the reference.
- p50 / p95 latency
- The median and 95th-percentile request time.
Clinical and data terms¶
- NEWS2
- The UK Royal College of Physicians' National Early Warning Score 2. It adds points for respiratory rate, oxygen saturation, supplemental oxygen, blood pressure, heart rate, level of consciousness and temperature. A total of 0 means routine observations; 1–4 is ward-level; 5–6, or any single parameter scoring 3, needs an urgent review; 7 or more is an emergency response. It is computed in code here.
- ACVPU
- The consciousness scale in NEWS2: Alert, new Confusion, responds to Voice, to Pain, Unresponsive. Anything other than Alert scores 3. Jev's new-confusion Noul can upgrade a charted A to C.
- SpO₂ scale 1 / 2
- NEWS2 oxygen-saturation scoring. Scale 2 is for patients with a lower target (88–92%), such as some COPD patients at risk of CO₂ retention.
- Escalation band
- This report's four response levels: routine → ward review → urgent review → emergency.
- Sepsis screen
- A structured check for infection causing organ dysfunction. It is prompted here when infection is suspected and NEWS2 is 5 or more.
- Delirium
- An acute, fluctuating change in attention and awareness. It is distinct from dementia, which is long-standing, but it often occurs on top of it.
- Medication reconciliation
- Comparing the medications a patient took before admission with the discharge plan, so every drug is deliberately continued, changed, withheld or stopped.
- Duplicate therapy / interaction
- Two drugs with the same ingredient or class taken together (for example two statins, or two paracetamol-containing products), or a combination known to be harmful (for example clarithromycin with simvastatin).
- Safeguarding
- Protecting a patient, or someone they care for, from harm, neglect or self-harm.
- Melaena
- Black, tarry stool from upper gastrointestinal bleeding. Black stool on iron tablets is expected and harmless.
- STEMI / NSTEMI / PCI
- Heart attack with or without ST-segment elevation on the ECG, and percutaneous coronary intervention (a stent).
- GTN
- Glyceryl trinitrate (nitroglycerin) spray for angina.
- FHIR R4 / R5
- HL7 Fast Healthcare Interoperability Resources, the standard for exchanging health records as typed JSON resources. R4 (4.0.1) is the most widely deployed; R5 (5.0.0) is the latest release.
- FHIR resources used here
Patient,Condition(diagnoses),AllergyIntolerance,MedicationRequest,Observation(vitals and labs),Encounter,DocumentReference(notes), andCommunication(patient messages).- US Core
- The US implementation guide that constrains FHIR R4. Synthea's R4 exports conform to it.
- Synthea
- An open-source generator of realistic but entirely synthetic patient histories.