Verdict

Appearance

Providers ​

Choose an adapter below, set its environment variables in the process running Verdict, and copy its defaultJudge into your rubric. Remote clients configure credentials on the API server, not in the client. Never put secret values in YAML.

TypeSafe Jev ​

The default provider has a live-verified contract. Set TYPESAFE_API_KEY:

bash
export TYPESAFE_API_KEY="your-typesafe-key"
yaml
defaultJudge: { type: decision, provider: typesafe, model: jev-latest }

Cloudflare Clef-flash ​

For REST inference, set both account ID and a token with Workers AI Read permission:

bash
export CLOUDFLARE_ACCOUNT_ID="your-account-id"
export CLOUDFLARE_API_TOKEN="your-workers-ai-token"
yaml
defaultJudge:
  type: decision
  provider: cloudflare
  model: "@cf/cloudflare/clef-flash"

Clef-flash has a live-verified fixture. A Worker with an AI binding uses that binding instead of REST credentials; see Workers deployment.

System One-compatible endpoint ​

Use this adapter for a service implementing the System One request and response shape. Verify the endpoint yourself; compatibility is defined by the adapter, not by a live fixture for your service.

bash
export SYSTEMONE_BASE_URL="https://your-service.example/v1"
export SYSTEMONE_API_KEY="your-service-key"
yaml
defaultJudge: { type: decision, provider: systemone, model: jev-latest }

Verdict appends /systemone to the base URL.

OpenRouter and OpenAI: provisional adapters ​

These decision adapters are based on public write-ups, without a live-verified contract in this repository. Confirm endpoint availability and response compatibility before integrating them. These are separate from generative LLM providers with the same names.

bash
export OPENROUTER_API_KEY="your-openrouter-key"
yaml
defaultJudge: { type: decision, provider: openrouter, model: typesafe/jev-1.13 }
bash
export OPENAI_API_KEY="your-openai-key"
yaml
defaultJudge: { type: decision, provider: openai }

Mock: offline integration ​

No environment variables or network access are required:

yaml
defaultJudge: { type: decision, provider: mock }

Mock results test plumbing; they do not assess evidence quality.

Credentials and evidence budgets ​

Use apiKeyEnv: REVIEW_KEY on a judge to read a custom environment variable instead of its provider’s default. Set options.stateCharLimit to limit rendered evidence; truncation is recorded in rating.raw.stateTruncated. Use criterion inputs tags to avoid sending unrelated documents.

The verified contracts and fixtures record endpoint shapes and verification dates. For batching, probability mapping, and escalation, follow decision models.

Written feedback ​

Use type: llm when you need generated rationales and quotes. Its provider contract differs from type: decision; see LLM judge fields for credential names and options.

Next, copy a judge into the minimal rubric, validate it, and evaluate representative evidence before choosing a confidence threshold.