Decisions API

A decision policy is a saved question for Jev with INFRO's thresholds around it. Your application sends only the facts (state); the question, what each outcome means, the confidence bands, pinned outcomes, the deadline and the fallback live in the policy, edited in the console under Decisions and versioned.

Decide

curl https://api.infro.io/v1/decisions/tool-gate \
  -H "Authorization: Bearer $INFRO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "state": {
      "request": "Tidy up my inbox",
      "tool": { "name": "delete_emails", "arguments": { "older_than_days": 30 } }
    }
  }'

statestring | object | arrayrequired

The facts the decision is about. At most 128 KB as JSON. Never stored.

deadline_msinteger

Overrides the policy's deadline for this call, 50–10000 ms.

metadataobject

Attribution and tags, as on every endpoint. Recorded on the row.

{
  "id": "req_…",
  "object": "decision",
  "policy": "tool-gate",
  "version": 3,
  "outcome": "ask",
  "label": "Should run only after the user confirms: it changes or deletes data, spends money, or contacts other people.",
  "band": "confirm",
  "confidence": 0.71,
  "source": "model",
  "probabilities": { "allow": 0.12, "ask": 0.81, "deny": 0.07 },
  "score": null,
  "latency_ms": 150,
  "model": "typesafe/jev-1.13",
  "usage": { "input_tokens": 412, "output_tokens": 0, "cost": 0.0000173, "price_source": "list" },
  "route": "primary",
  "route_name": null,
  "provider": "typesafe",
  "connection": "conn_…"
}

import os
import requests

answer = requests.post(
    "https://api.infro.io/v1/decisions/inbox-triage",
    headers={"Authorization": f"Bearer {os.environ['INFRO_API_KEY']}"},
    json={"state": {"from": "dana@customer.example", "subject": "Invoice before Friday", "body": "…"}},
    timeout=10,
).json()

if answer["band"] == "act":
    route_email(answer["outcome"])
elif answer["band"] == "confirm":
    ask_user_to_confirm(answer["outcome"])
else:
    send_to_a_person()

Pins

A pin decides by rule before any model call: when a dot path into the state (tool.name) is, is one of, starts with, ends with or contains a value, the outcome is fixed. The first matching pin answers with source: "pin", confidence 1 and band act, costs nothing and needs no connection. Use pins for the cases that must never depend on a classifier — a tool that may never run.

Deadline and fallback

A policy can set a deadline. When Jev has not answered by then — or when every connection fails — the policy's fallback answers with source: "fallback", the band the policy chose for it and a fallback_reason. A missed deadline is never counted against the health of your connection. With no fallback, a missed deadline is 408 decision_deadline and a provider failure is the usual chain-exhausted error.

Batches

curl https://api.infro.io/v1/decisions/ticket-labels/batch \
  -H "Authorization: Bearer $INFRO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "concurrency": 8,
    "items": [
      { "id": "t-1001", "state": { "title": "Charged twice", "body": "…" } },
      { "id": "t-1002", "state": { "title": "Export hangs", "body": "…" } }
    ]
  }'

Up to 500 items per request, decided on one of your connections at the chosen concurrency (default 8). Each result carries its index and your id. An item your provider rejects is that item's error; the others are still decided. summary counts outcomes, bands and sources (and the mean score for a score question), so a large table is labelled as a series of batches whose summaries add up. A batch is one row in the request log.

List policies

GET /v1/decisions lists the enabled policies the key may call, with each one's question type, outcomes and what they mean, bands, pin count, fallback and deadline. A policy limited to a project is listed only for that project's keys.

Eval monitors

A policy with an eval monitor scores a share of your served chat completions in the background, after the answer has been returned: the policy sees {prompt, response} (the latest user message and the model's answer). Each score is its own call to the policy's decision model, on your own account with its provider, recorded under the request it judged; the policy's console page shows the mean score per served model. Answers served from INFRO's response cache are not scored again.

Errors