The guarded decision layer for AI agents
Your agents make judgment calls constantly — route this, flag that, is this risky. Decision Models turn those calls into typed, calibrated, auditable decisions, powered by JEV and run through the same guarded Anoman gateway as every chat call.
The problem
Fuzzy judgments, forced through a chat model
Most agent pipelines make dozens of small judgment calls per request — which team should this ticket go to, is this content risky, should this tool call run without a human. The default tool for that job is a chat completion: an LLM generating free text that your code then has to parse, hope stays on-schema, and re-prompt when it doesn't. It's slow, it's unstructured, and a model that can say anything can also say the wrong thing outside the shape you asked for.
What it is
Typed questions in, calibrated answers out
A Decision Models call takes a state — the thing being judged — and a set of typed questions, and returns a typed answer for each one with a calibrated probability attached. There's no free text generated at any point, so there's nothing for the model to hallucinate outside the schema you defined.
noulYes / no
A binary judgment with a calibrated probability — is this urgent, is this spam, does this need review.
choiceClassify
Pick one label from a fixed set you define — route a ticket, categorize content — with a confidence score per option.
scoreRate
A graded rating along a scale you define — tone, severity, quality — with the full probability distribution across grades.
Billed on input tokens only — output is free — and a typical call resolves in well under a second.
Use cases
Where typed decisions replace a chat call
Ticket & intent routing
Classify an inbound message into a team or workflow branch — before a human, or an expensive model, ever sees it.
Content & severity classification
Flag risky, sensitive, or policy-relevant content with a calibrated score, not a guess.
Scoring an LLM's own output
Rate a chat completion's quality, tone, or groundedness before it reaches a user — a cheap judge in front of an expensive generator.
Agent tool-call risk gating
Decide whether a proposed tool call is safe to run automatically, or needs a human in the loop.
Guarded, honestly
What actually happens to your data
Only state is redacted. PII detected in state is masked before it leaves Anoman. Your questions and instructions are sent to JEV exactly as authored and are not scanned or redacted — don't put secrets or sensitive identifiers in your question text.
US routing. Unlike the rest of the Anoman gateway (single-region Jakarta), Decision Models route your request to JEV's API in the United States. If your content must stay in-region, don't send it to this endpoint.
Injection scoring is monitor-only. A judge legitimately has to evaluate adversarial or attack-shaped text as part of its job — so prompt injection on state is scored and logged, but this endpoint never blocks on it.
Decisions are probabilistic. Every answer is a calibrated probability, not a certainty. Review important decisions before acting on them, especially where the returned confidence is low.
Give your agents a typed decision layer
Read the full API reference, or start calling it against your own traffic — the same key you already use for chat completions works here too.