ClearRun

Honesty scoring for AI responses. ClearRun evaluates any AI output and returns a plain-language honesty verdict through a small open SDK, model adapters, and a CLI. A verdict includes a score, a grade, a failure breakdown, and a signed, verifiable record.

What ClearRun measures: confidence-justification alignment, whether an AI's expressed confidence is backed by its reasoning. It does not fact-check. A high score means the response's confidence matches its justification, not that it is factually true. Verified factual accuracy is on the roadmap.

Install

All six @clearrun packages are on the npm registry. @clearrun/sdk and @clearrun/cli are at 0.1.1; @clearrun/core-types, @clearrun/scoring, @clearrun/security and @clearrun/vendors are at 0.1.0. See CHANGELOG.md for what changed in 0.1.1.

npm install @clearrun/core-types @clearrun/sdk @clearrun/scoring @clearrun/security @clearrun/vendors
npm install -g @clearrun/cli

To work from a clone of this repository instead:

git clone https://github.com/adepoco/clearrun.git
cd clearrun
npm install

Then import @clearrun/sdk from the workspace.

If this repository lives at a different URL, clone that URL instead. Tag v0.1.0-oss is the release snapshot.

Quickstart

import { createKernelClient } from '@clearrun/sdk';

const client = createKernelClient({
  hosted: { apiKey: process.env.CLEARRUN_API_KEY }, // required; get a beta key at api.clearrun.net
});

const { verdict } = await client.evaluate({
  responseText: 'The Eiffel Tower is in Berlin.',
});

console.log(verdict.score, verdict.grade, verdict.failureCategories);

No API key configured? Pass { forceMock: true } for local development. The mock returns schema-valid verdicts, never real scores.

What you get back

A Verdict (see schema/verdict.schema.json and docs/CONTRACT.md): score (0-100), grade, risk, failureCategories[], confidenceInterval, componentSignals[], plain-language summary, and an Ed25519 signature you can verify locally, or via GET https://api.clearrun.net/api/verify/{verdictId}.

Open core

• MIT (this repo): the SDK, model adapters (OpenAI, Anthropic, Google), CLI, verdict types, and schema. Clone it, embed it, build on it.
• Hosted, closed (beta): the scoring kernel runs as a hosted service at api.clearrun.net. It provides governance protocols, honesty scoring, and signed attestation. The SDK calls it; the scoring logic is not in this repo by design.

Roadmap

• Now: confidence-justification alignment scoring (this release). Calibrated for English; scores on non-English output are outside today's validated scope.
• Next: multilingual scoring (language-specific confidence and hedging cues), a verified factual-accuracy layer, richer audit, and enterprise governance exports.

License

MIT for everything in this repository. The hosted scoring kernel is licensed separately (BSL-1.1). See LICENSE.