# JoinAI dependency-free LLM evaluation harness

This small reference harness runs deterministic checks over JSONL cases using only the Python standard library.

```bash
python3 harness.py sample.jsonl --report report.json --min-pass-rate 1.0
python3 -m unittest -v test_harness.py
```

Replace `replay_candidate` with an adapter for your application. Keep credentials outside the dataset and do not commit sensitive production prompts or outputs.

Each JSONL object requires `id`; optional fields include `input`, `candidate_output`, `slice`, and `expected`. Supported expectations are `exact`, `contains`, `forbidden`, and `pattern`.

The harness is intentionally small. Add application-specific graders, version envelopes, slice reports, retry policy, and redaction controls before using it as a release gate.