verdict/decision
API keysRequest access

API / QUICKSTART

One endpoint. Typed decisions.

Send relevant state and one or more bounded questions. The API returns probabilities and a typed assessment for each question, together with a model and execution receipt.

Connect to the gateway

Use POST https://api.verdictdecision.com/v1/decisions. Supply a vd_test_ or vd_live_ key in the Authorization header. Keys are not open to customers yet; see API access. A Verdict (vrd_) key is not a Decision key.

Set VERDICT_DECISION_KEY in your environment. The examples use HTTPS and standard libraries, without a separate Verdict SDK.

curl https://api.verdictdecision.com/v1/decisions \
  -H "Authorization: Bearer $VERDICT_DECISION_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: $(uuidgen)" \
  --data '{
  "model": "decider-4b-v2",
  "state": {
    "issue": "The checkout page displays an incorrect total after applying a coupon."
  },
  "questions": {
    "result": {
      "type": "choice",
      "instructions": "Choose the team best suited to investigate this issue.",
      "criteria": {
        "frontend": "UI rendering and browser behavior",
        "billing": "Pricing, discounts, and payment calculations",
        "infrastructure": "Deployment and service availability"
      }
    }
  }
}'

Request fields

Three question types

Proposition: a yes/no assessment. The answer includes a boolean decision, a probability of true, and abstained.

Choice: define 2–255 explicit criteria (2–64 on the Laya models). The answer includes the selected option, probabilities for the options, confidence, and abstention. When using cyclic or tournament strategies, choice answers include game_theory receipts reporting position bias cancellation, margin, and Condorcet consistency.

Score: define 2–10 ordered levels. The answer includes the selected numeric level, an expected score, and probabilities over the levels.

Game-theoretic decision strategies

For high-stakes choices such as patch selection, bug triage, or multi-candidate code generation, naive repeated sampling fails because errors between runs of the same model correlate heavily. The Decision API provides two mathematically rigorous social choice mechanisms:

Read the technical benchmark study for complete local measurements, Wilson confidence intervals, and raw receipts.

Recipes

Receipts and retries

Read results from answers[question_id]. The response also identifies the resolved model, serving venue, artifact hash, runtime fingerprint, and settlement state. Early-access receipts currently report not_billed_pre_t4.

Supply an Idempotency-Key and keep the same key and body for a manual retry. Completed successful responses are cached in gateway memory; restarts clear that cache. Failed or interrupted calls can have an unknown upstream outcome. They are not automatically safe to re-execute.

Latency and limits

Each model normally runs on the GPU named above. When that GPU is scarce, the platform may serve it from a second GPU type, so latency can differ. The worker scales to zero. An idle worker can need a cold start, so a deadline can expire before a result is ready. There is no automatic retry in the playground. The direct API accepts up to 32 questions (16 on the Laya models) and 256 KiB per request. The cyclic and tournament strategies expand each choice into K or K*(K-1) worker questions, and a request that expands to more than 1,000 is refused with 422 expansion_too_large; the playground accepts up to four questions.

Errors

Failures use {"error": {"code", "message"}} and an HTTP status. Retry only where the table says so.

Available models & authority

The Decision API serves eight open-weight models under their upstream names. GET /v1/models reports each model's provenance: author, upstream repository and revision, license, and whether Verdict modified the weights.

Specify the model in the request body. Initial abstention thresholds are not yet fitted. A probability is an assessment; it does not grant tool permission or establish that an execution succeeded.

View the live model registry · Check gateway status · Try the playground

Licenses and attribution

Verdict Decision serves open-weight models. Each is listed under the name its authors publish it as, with the upstream repository, pinned revision and license shown below and in GET /v1/models. Names such as Decider, Laya and Qwen belong to their respective authors and identify the upstream models only. They are not Verdict Decision product names.

Licensed under the Apache License, Version 2.0. Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" basis, without warranties or conditions of any kind.