verdict/decision
API keysRequest access
Decision API · one forward pass

Ask a bounded question. Get a typed verdict.

Verdict Decision takes state and a question and returns a yes/no probability, a choice between options, or a score. It doesn't generate text, so you have nothing to parse and can branch on the answer straight away.

Request API accessTry the playground
HTTPS + JSON · Python and TypeScript examples below
request
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"
      }
    }
  }
}'
responseIllustrative example

Which team should handle this issue?

frontend
0.24
billing
0.68
infrastructure
0.08
{
  "type": "choice",
  "decision": "billing",
  "confidence": 0.68,
  "certainty": 0.44,
  "abstained": false,
  "probabilities": {
    "frontend": 0.24,
    "billing": 0.68,
    "infrastructure": 0.08
  }
}

Example probabilities show the response shape. They are not measured performance.

Edit request JSON

Your key stays in this tab’s memory and is forwarded securely to the Decision gateway. Cold starts can take longer. Request access.

01 — why decide

Stop parsing prose to make a choice.

a generator
“Looking at the command, it appears to delete the build folder, which is generally safe, although it could be risky if the path is wrong, so I would probably suggest…”
→ regex it, retry it, hope it parses
a decision model
{ "decision": true, "probability": 0.94 }
→ assess, then apply your permission policy
02 — the api

Three answer types. One endpoint.

Every call goes to POST /v1/decisions. You set the type, and the response comes back in the matching schema every time.

Binary / Proposition
A yes/no probability. Use it to assess checks and gates; your permission policy stays in charge.
{ decision: boolean
  probability: number
  abstained: boolean }
Choice
Pick one of the options you list and get a probability for each. Use it for routing, labelling and ranking.
{ decision: Option
  probabilities: Record<Option, number>
  abstained: boolean }
Score
A level on a scale you define. Use it for risk, urgency, quality and priority.
{ decision: number
  expected_score: number
  probabilities: Record<Level, number> }
03 — use cases

Every place your code asks “which one?”

Replace a prompt-and-parse step with a bounded judgment. Measure latency, cost, and quality on your own workload.

Model routing

Which model and effort level fit this request?

choice

Tool assessment

Does this action need human review?

proposition

Triage

How urgent is this ticket, 1–5?

score

Classification

Which team owns this bug?

choice

Guardrails

Does this reply expose a secret?

proposition
research · local pilot

Real coding decisions, measured locally.

In an empirical pilot on 90 functional code discrimination tasks, rotating options selected the passing implementation in 89 of 90 tasks, versus 85 for a single pass.

SINGLE PASS · 1 FORWARD
85/90 (94.4%)
98 ms local median · 151 ms p95 · Holdout ECE 5.6%
CYCLIC ROTATED · 3 FORWARDSHIGHEST ACCURACY
89/90 (98.9%)
291 ms local median · 454 ms p95 · Cancels position bias algebraically
See results and method →Measured on Apple M5 Max with resident 4B MLX model. Hosted API performance was not evaluated.
04 — deploy

Run it next to your data, or call it from anywhere.

Local
Local inference is part of the launch plan, on compatible hardware. The free local venue keeps decision inputs on your machine.

Local setup and hardware requirements are being finalized.

Verdict Cloud
Typed decisions over HTTPS across multiple open-source profiles: 4.0B Decider for deep multi-model adjudication, plus 421M Laya Typed (~22 ms) and 322M Laya Multilingual for high-throughput edge triage.
https://api.verdictdecision.com/v1/decisions
observe
Research direction: use consented internal outcomes to understand which task configurations work.
label
Tests and Smart Harness validation provide outcome labels. A probability never replaces verification.
train
Future policies can learn from verified outcomes. Customer inputs are not used for training without opt-in consent.
05 — pricing

Pay per decision.

local
$0
Launch plan: free inference on compatible hardware. Local setup is being finalized.
Download Verdict
cloud● hosted
Early access
API access is currently provisioned by request. Public credit pricing and self-service key issuance are not open yet.
Request access
enterprise
Custom
Discuss capacity, private deployment, and your integration requirements with the Verdict team.
Talk to us

Your code has questions. Give it verdicts.

Request API accessRead the docs