one of a setDocs ↗
Choice
Jev picks one option from a list you define and gives the odds for every option, plus a confidence.
Decisions, not prose.
Jev doesn't write text. You send it some content and a few typed questions; it sends back answers your code can use straight away. Your code owns the rules — Jev supplies the judgement.
yes / no probabilityDocs ↗
Noul
Jev returns how likely the answer is yes, from 0 to 1. Ask one per condition when several can be true.
graded levelDocs ↗
Score
Jev rates something along levels you describe; the answer can land between levels so rankings stay smooth.
Five patterns behind every idea
pick one of a set
Docs ↗Choice Router
How: a Jev `choice` question that routes each incoming item to one handler/category from a fixed option map, plus branch-specific questions asked in the same request
UI: a paste box that shows the chosen option, the full probability bar chart, and a 'needs review' badge when confidence is low
yes/no probability
Docs ↗Noul Check
How: several independent Jev `noul` questions (one per condition) asked together over the same state, combined by explicit code rules
UI: a checklist of conditions, each with a 0-1 probability meter and a pass / review / fail colour from code-owned thresholds
graded ranking
Docs ↗Score Ranker
How: comparable Jev `score` questions per candidate on 3-4 named dimensions, normalised with score/(levels-1) and combined with user-controlled weights in code
UI: a ranked list with weight sliders that re-rank instantly without calling Jev again
pick the real value
Docs ↗Select, Don't Generate
How: code extracts candidate spans/values from the source, then a Jev `choice` (with a 'none' option) selects the intended one so the app copies real text instead of generating it
UI: the source text with candidate spans highlighted and the selected one locked in
check then hand off
Docs ↗Verify & Escalate
How: a Jev `noul` checks each claim/field against its evidence; low or mid probabilities are escalated to a human review queue instead of auto-acting
UI: claim cards marked supported / unsupported / needs review, with a review queue column
One request, many answers
POST https://api.typesafe.ai/v1/systemone
Authorization: Bearer $TYPESAFE_API_KEY
{
"model": "jev-latest",
"state": { "message": "Our venue cancelled — need a space by Thursday!" },
"questions": {
"team": { "type": "choice", "instructions": "Which crew handles `message`?",
"criteria": { "booking": "Venues and dates", "payments": "Fees", "none": "Nothing needed" } },
"urgent": { "type": "noul", "instructions": "Does `message` need a reply within 24h?" }
}
}
→ answers.team.choice = "booking", answers.urgent.noul = 0.93Keep TYPESAFE_API_KEY on the server. Questions in one request run in parallel and can't see each other's answers.