Page 01 · Fit test
What can I build with Jev?
Use Jev where software needs a bounded judgment more often than it needs a paragraph.
Model shape: state plus typed questions in; a Choice, Score, or Noul answer with probabilities out. The answer space is defined before the call, so the output can be consumed without free-form parsing.
Choice
Route or select
Pick one of your labels: support queue, document type, model tier, tool, lead bucket, or next workflow step.
Best when the options are known and the code can reject low confidence.
Score
Prioritise or rank
Place a state on an ordered rubric: urgency, risk, quality, fit, or review level.
Best when you can define what each level means and validate it on real examples.
Noul
Gate or escalate
Ask whether a condition is true: escalate, allow, complete, safe, or worth a human review.
Best when the probability feeds a conservative threshold in your code.
Patterns that fit the boundary
| Pattern | Jev decides | Your system still owns |
|---|---|---|
| Support triage | Topic, urgency, escalation probability | Queue permissions, reply text, SLA policy, human handoff |
| Model routing | Complexity or risk bucket | Which provider to call, budgets, retries, and the final response |
| Tool-call guardrail | Whether a proposed action looks risky | Allowlist, authentication, confirmation, and execution |
| Moderation or review | Policy label and review score | Appeal path, evidence, enforcement, and audit record |
| Data extraction | Which declared field or category fits | Schema checks, missing values, storage, and correction flow |
A safe first spike
- Pick one repeated branch. Start with a decision that currently costs a slow LLM call or brittle rules—not the whole product.
- Write the answer space. Name the options or 2–10 score levels. Define what a “yes” means for Noul.
- Measure on labelled examples. Record accuracy, calibration, latency, state size, and the cost of a wrong decision.
- Keep a fallback. Route uncertain or high-impact cases to a rule, larger model, or person until the evidence earns automation.
Not a replacement for: writing, code generation, summarisation, open-ended conversation, arithmetic, or a policy engine that must be deterministic. “Cannot hallucinate” means Jev stays inside your declared output space; it can still choose the wrong valid option.
Sources checked 20 Sep 2026
- TypeSafe launch post — official use-case framing and vendor-reported claims.
- TypeSafe primitives — Choice, Score, and Noul contract.
- TypeSafe patterns — workflow patterns and code-led control.
- TypeSafe playground router — an official demo that explicitly mocks execution.