Jev Project Classifier

Jev classifies. GitHub Stars rank. Humans decide what gets published.

This is an independent community tool for classifying Jev-related GitHub projects, System One ecosystem work, and real-world use cases. It supports the OpenRouter Jev Decisions API used by Jev AI Dev and TypeSafe's official @typesafe-ai/sdk.

GitHub Star ranking is deliberately simple and reproducible: current Star count in descending order. Jev does not score project popularity or decide the final list order.

This project is not an official TypeSafe AI release. Official Jev information is available at typesafe.ai.

What the first version does

• Collects named repositories or GitHub search results, README text, metadata, current Stars, and latest commit time.
• Applies one of three versioned Jev rubrics: project, system-one, or use-case.
• Produces structured JSON with suggested tags and explicit review flags.
• Produces a Markdown checklist for maintainers.
• Marks every result pending_human_review.

The CLI does not publish, commit, push, open pull requests, edit Jev AI Radar, or update a website.

Validated on a real editorial workflow

On September 23, 2026, Jev AI Dev ran this tool against 15 items from its live Projects, System One, and User Cases collections. Across the 10 Project and User Case items with scenario baselines, Jev's primary category matched one of the existing human labels in 10 of 10 cases. The sample is small and does not establish general accuracy. See the validation note for the method, findings, and limitations.

The repository includes a maintainer adapter for a local checkout of the jevai.dev source. It parses the expected JSON and JSON-like TypeScript data declarations without importing or executing the source files. Other users can provide the same optional metadata.humanBaseline.scenarios field in their own candidate JSON and run the generic evaluate command.

npm run cli -- import-jevai-dev \
  --site-root ../jevai \
  --profile project \
  --limit 5 \
  --output data/pilot-projects.json

For project and system-one, the pilot fetches current repository evidence from GitHub. For use-case, it imports the source URL and the evidence notes currently used by the live page. After classification, compare suggested scenario tags with the current human labels:

npm run cli -- evaluate \
  --input data/pilot-decisions.json \
  --output data/pilot-evaluation.json

The evaluation reports primary-category agreement, precision, recall, exact matches, missing tags, and extra tags. Agreement is a workflow signal rather than proof that either label set is correct; a maintainer reviews every disagreement.

Setup

Requires Node.js 20 or newer.

npm install
cp .env.example .env
chmod 600 .env

The default local configuration uses the same OpenRouter route as jevai.dev: OPENROUTER_API_KEY with model typesafe/jev-1.13. Set JEV_PROVIDER=typesafe and TYPESAFE_API_KEY only when using TypeSafe's direct API. GITHUB_TOKEN is optional but raises GitHub API rate limits. The tool never writes any key to its output.

The CLI loads a local .env automatically. That file is excluded by .gitignore.

Data sent to the model provider

The classify command sends each candidate's source metadata, summary, supplied content, and GitHub evidence to the configured OpenRouter or TypeSafe API. Human baseline labels under metadata are not sent. Only classify public material or content you are authorized to share with that provider, and review the provider's privacy and retention settings before processing sensitive data.

Workflow

1. Collect known repositories

npm run cli -- collect \
  --repo typesafe-ai/typesafe-sdk-js \
  --output data/candidates.json

Or discover candidates from a GitHub query:

npm run cli -- discover \
  --query "jev in:name,description,readme" \
  --limit 20 \
  --output data/candidates.json

Discovery is candidate generation, not inclusion. Search results still require Jev classification and human review.

2. Create a deterministic Star ranking

npm run cli -- rank-stars \
  --input data/candidates.json \
  --output data/ranked.json

3. Classify and tag with Jev

npm run cli -- classify \
  --profile project \
  --input data/candidates.json \
  --output data/decisions.json

For community models or use cases, choose --profile system-one or --profile use-case. Use-case candidates can begin with examples/use-cases.json.

4. Review before any publication

npm run cli -- review \
  --input data/decisions.json \
  --output data/review.md

The review queue asks a maintainer to verify source attribution, relationship, licensing, identity, claims, and tags. Accepted records can later be transferred to Jev AI Radar through a separate, deliberate editorial process.

Output contract

Each classification preserves the complete candidate and adds:

• profile and rubricVersion
• starRank, calculated independently from model output
• raw normalized Jev answers
• suggestedTags
• primaryTag, taken from the single-choice category question
• secondaryTagCandidates, kept with probabilities for human review
• reviewFlags
• recommendation
• reviewStatus: "pending_human_review"
• model name, token usage, and classification time

Read the architecture and rubric definitions for the publication boundary and interpretation rules.

Candidate and result fields are documented in the data format.

Development

npm run check

The example records use placeholder URLs and are not real directory recommendations.