Documented systems · 2026
Small, typed questions let an application use model answers in ordinary control flow. Explore the public API through a support-routing example with an explicit review policy.
TypeSafe AI · Public documentation
START WITH THE IDEA
Jev’s public interface lets you ask named, typed questions about a state. Choice returns a category with probabilities and confidence; Noul returns a scalar. Your application can validate those answers and compose them with ordinary code into a decision.
THE SIGNIFICANCE
A named question and a declared answer type give software a clearer contract than parsing an unconstrained paragraph. Structured-output general models can also provide such contracts; the comparison here makes that shared requirement visible.
Request and response examples; Hopper comparison ↗A proposed category need not trigger an action. The application can apply a threshold and ask for review. Treat confidence as a documented model signal, then evaluate its usefulness on your own labeled cases; it is not automatically a calibrated probability of correctness.
Confidence documentation ↗The useful material here is the public API and its compositional patterns. The browser demonstrates policy using authored fixtures. The downloadable Node client implements that documented request contract; no live-provider result or private model architecture is claimed.
Patterns; Hopper integration scope ↗Choose a request, inspect the proposed department and change the threshold required for automatic routing. Urgency remains a separate decision.
Compare the approaches. Three executable paths on one request: keyword rules, validated structured JSON and a Jev-shaped typed answer. Both model paths use the same authored values to avoid inventing a performance difference.
From the implementation repository after cloning:
node packages/jev/cli.mjs mixedSource & instructions ↗
Authored response fixtures compare keyword rules, validated structured JSON and a Jev-shaped response. No provider calls, credentials or model performance measurements in the browser.
CHOOSE THE RIGHT TOOL
Hopper’s application guidance. Compare on your own task before adopting a more complex approach.
Ask small typed questions, validate the answers and keep the decision policy in code. Evaluate the real API on representative cases first.
The lab demonstrates the interface and policy; the Node client is available for your own account and evaluation.
Use ordinary code, a parser or a lookup when the contract is deterministic.
Compare the keyword baseline, then inspect cases where keywords miss context before adding a model.
Keep it if it satisfies your validation, quality, latency and cost requirements. The two model-shaped paths here deliberately use identical answers.
Measure on the same labeled cases before switching providers; the fixture lab is not a model ranking.
A confidence value alone is not sufficient authorization or proof of correctness. Keep review and policy controls in the owning application.
Test calibration and failure modes for your task, and separate a proposed decision from executing it.
import { evaluateJev } from './packages/jev/client.mjs';
import { routeAnswers } from './packages/jev/index.mjs';
// Node only; obtain apiKey from a private file.
const response = await evaluateJev(ticket, { apiKey });
const decision = routeAnswers(response.answers, { threshold: 0.8 });
console.log(decision); // proposal, not an actionFree open-source code · GitHub ↗Use typed questions for triage, extraction and other bounded decisions. Validate answers at the application boundary, evaluate thresholds against labeled examples from your workflow, and keep a human review path.
This is an independent integration example based on public documentation. Offline fixtures are authored, not recorded model results. The Node client has contract tests with a stub; it has not been qualified with live Jev access. Neither speed nor accuracy is benchmarked here.
node packages/jev/cli.mjs mixed
THE ORIGINAL SOURCE
Our explanation and runnable example are here. The original authors maintain the source and documentation linked below.
TypeSafe AI documentation ↗Typed model answers can feed an execution loop. Our examples keep policy and verification explicit; this is an editorial connection, not a documented dependency.
The endpoint, authorization and named-question format used by the downloadable client.
Request and response contract · Source documentationDefines categories, probabilities and confidence.
Choice answer fields · Source documentationExplains the confidence signal. It is not automatically a measured probability of correctness on your task.
Interpreting confidence · Source documentationSmall questions and ordinary code compose into larger decisions.
Composing questions · Source documentationDocumentation reviewed 2026-09-20. Public documentation reviewed September 2026.
Original source ↗ · Documentation ↗
Explanations, diagrams and educational code by Hopper Labs. Code: MIT. Original publications and documentation remain with their authors. These are independent educational examples, not endorsements. Any recorded evaluation is described separately with its exact scope.