Yes/no, choice and score decisions for agents, about a tenth of a cent each

Two ways to call it

The model reports a probability for a predicate and a distribution for a choice or a score, rounded to two places. We have not measured how well those probabilities line up with real outcomes, so we do not call them calibrated. Treat them as the model's stated confidence and pick a threshold from your own labelled examples.

Here is why that matters. Ask the support router — billing, shipping, technical — about “The mitochondria is the powerhouse of the cell.” It answers technical at 0.94, three times out of three. The answer is wrong and the number does not warn you: a choice question always picks one of the values you offered.

The fix is in the question, not the threshold. Add a value for “none of these” and the same sentence comes back none at 0.99; a predicate asking whether it is a support request at all returns 0.0. Give every choice somewhere to put the case that fits nothing.

One call: the evidence, the questions, the answers

curl https://api.blockrun.ai/v1/decisions \
  -H "authorization: Bearer $BLOCKRUN_API_KEY" \
  -H 'content-type: application/json' \
  -d '{
    "model": "gpt-6-luna",
    "input": "Help! My payouts have been failing for 3 days.",
    "questions": [
      { "type": "predicate", "name": "is_urgent",
        "instructions": "Does this convey urgency?" },
      { "type": "choice", "name": "department",
        "instructions": "Which team should handle this?",
        "choices": [ { "value": "billing",   "description": "Payments, refunds" },
                     { "value": "technical", "description": "Bugs, outages" },
                     { "value": "none",      "description": "Fits neither" } ] },
      { "type": "score", "name": "frustration",
        "instructions": "How frustrated is the customer?",
        "levels": [ { "label": "calm",       "description": "No frustration" },
                    { "label": "frustrated", "description": "Clearly annoyed" },
                    { "label": "very angry", "description": "Hostile" } ] }
    ]
  }'

# Billed from your key's credit. With the OpenAI SDK:
#   OpenAI(base_url="https://api.blockrun.ai/v1", api_key=BLOCKRUN_API_KEY)

The response is { model, answers: [...], usage }, one answer per question in the order you asked. The key comes from user.blockrun.ai . Without a key, send the same body to /v1/decisions on blockrun.ai and pay the 402 from a wallet.

Where a judgment sits in a real workflow

1. Gather the evidence: the customer's message, the transactions it refers to, the policy that governs it.
2. Ask the independent questions together in one call.
3. Combine the answers with deterministic checks your code already has: the amount, the account age, prior refunds.
4. Route. Act automatically where the answer is clear and the checks agree; send the rest to a person or a reasoning model.

For the slice that needs deliberation, a reasoning model is one endpoint over, on the same key.

OpenJev is retired

Until 2026-10-07 this endpoint was answered by OpenJev, an open-source natural-language-inference model we hosted ourselves. It is gone. Its numbers were entailment shares rather than probabilities, it read sarcasm as praise, and its first call after a quiet spell took about half a minute.

If you integrated against /v1/decide, it still answers: the old noul / choice / score request is translated to this one. New code should send the Decisions shape above. Jev, TypeSafe's model, is a different product that we never served; OpenJev vs Jev keeps the history.

What this is not

It does not write, summarise, or explain. If what you want is a typed value out of a chat model instead, that is a different problem with its own traps — what response_format actually does per provider has the measurements. If the job is sorting text into labels you can name up front, the text classification API is this endpoint written up for that one job. For prose, the chat models are next door.

It has no memory between calls and judges what you send rather than looking anything up. If the judgment needs a fact from the web, fetch it with search and pass it in as evidence.