Jev
OpenRouter decisions · ~typesafe/jev-latest
idle
Run a decision to see this provider’s typed output.
The same triage question, answered two ways.
Paste a customer-support message and route it once through a typed decision API and once through a language model, then compare what each one actually returns.
Baseline chat model
chat completions · openai/gpt-4.1-mini
idle
Run a decision to see this provider’s typed output.
Decisions, not prose
Jev returns a typed decision. The question schema is part of the request, so the answer is a value inside a declared space: one of five options, a number on a 0–3 scale, a 0–1 truth value. “No hallucinations” means exactly this — the output cannot escape the schema. It does not mean the decision is always right. Jev can route a ticket to the wrong team; it simply cannot invent a team that does not exist.
The language model generates text. Asking for JSON makes that text more likely to be parseable, but the result is still a string that must be parsed, validated, and sometimes rejected. This app records those steps rather than papering over them, which is the whole difference: one path produces a decision, the other produces a document about a decision.
Any accuracy or latency figures published by either vendor are vendor claims. This app measures only its own round-trip latency and schema outcomes for the single ticket you submitted; it is not a benchmark.
Choice
One option out of a closed, named set. The answer is a member of the set you declared — it cannot be a new label, a synonym, or a sentence. You also get a probability for every option.
Score
An ordered scale defined by your criteria — here 0 to 3. The returned value may be fractional, sitting between two rungs, with a probability mass over each rung.
Noul
A single truth value between 0 and 1 for a statement you assert. 0.93 means 'very likely true', not 'true'. It is a likelihood, never a yes/no verdict.
Confidence is not a probability. For Choice and Score, the probability distribution says how the mass is spread across options; the separate confidence value says how settled the decision is overall. A choice can hold 60% of the mass and still carry low confidence when the remainder is split across close alternatives.