Local Harness

Local Harness.

Ollama. Candidates. Verifiers. Disagreement.

Run

cargo run -- init
cargo run -- models
cargo run -- run cases/
cargo run -- report runs/<id>

Build

cargo build --release
./target/release/local-harness models

Config

{
  "endpoint": "http://127.0.0.1:11434",
  "generator": "qwen2.5-coder:7b",
  "verifiers": ["qwen2.5-coder:7b", "ilivo-local:latest", "qwen2.5-coder:7b"],
  "candidates": 4,
  "temperature": 0.7,
  "verifierTemperature": 0.1,
  "acceptScore": 80
}

Case

{
  "name": "short answer",
  "system": "Answer shortly.",
  "prompt": "2+2?",
  "rubric": "Correct and concise.",
  "expect": {
    "contains": ["4"],
    "regex": []
  },
  "tags": ["smoke"]
}

Use

• init
• models
• run
• report
• --json
• 3 verifier slots
• verifier disagreement
• JSON + Markdown reports

Later

• AB-MCTS
• SPELL loop
• latent plan search
• OpenAI-compatible provider
• GUI
• human label UI

Weight

About 1.5MB on macOS release build.

License

MIT.