Documentation

A bounded check, with evidence you can inspect.

Start with the contract, configure its authority, validate it, then verify a claim or observe an action transition.

Test a sample

Start with a few public cases before planning a larger evaluation. Keep reference answers out of model prompts.

Sample run instructions →

Evaluate your AI

Define the task and evidence rules, run a proof evaluation, then review failures and separate validation.

The customer workflow →

Integrate Runtime

Use the self-hosted API for supported claims and independent action reads. Managed onboarding is not available yet.

API and SDK →

Run a public sample

The public sample is for exploration, not untouched validation or clinical decision-making. Inspect source dates and limitations before interpreting results.

For a self-contained medical starter with 25 source-backed cases, download drug-indication-starter.zip. Its README explains prompt isolation, offline grading, optional model endpoints, source scope and conservative result interpretation.

  1. Or download groundtruth-sample.jsonl and the legacy evaluation runner. Review the script before executing it.
  2. Use Node.js with built-in fetch and configure any model provider credentials through environment variables, never source files.
  3. Start with five cases using the command below. Replace YOUR_MODEL_ID with a model available through your configured endpoint. Provider calls may incur charges.
  4. Review the output and source evidence. The legacy runner uses an LLM judge for provisional grading; this is not an authoritative Verification Engine receipt.
node eval-runner.mjs --provider openai --model YOUR_MODEL_ID --limit 5 groundtruth-sample.jsonl

For this command, use the file with expected_answer and grading_context. The separate train.jsonl uses verified_answer and verification_summary; it is not a drop-in replacement. Only send each question to your model—not reference answers, source explanations or prior responses.

For a no-code review, inspect the drug-indication sample. MED-RT relationships describe a bounded source view, not a complete prescribing label.

API essentials

Authenticate with a bearer API key. POST requests require an Idempotency-Key header. Reusing a key with different input returns a conflict. Results and receipts are tenant-scoped.

Claims return VERIFIED, CONTRADICTED, UNSUPPORTED or INDETERMINATE. Actions additionally distinguish NOT_VERIFIED, PARTIALLY_VERIFIED and VERIFIED_WITH_SIDE_EFFECTS. Only a clean VERIFIED result passes the default policy gate.

Full request examples, migrations and setup instructions live in the repository documentation. An isolated Cloudflare staging API has exercised the publication-year customer workflow against live NCBI evidence. It is operator-accessed staging, not a public production service. Other sources and action readers need separate managed validation.

Developer entry points →