Metadata-Version: 2.4
Name: trinity-claim-trace
Version: 0.1.0
Summary: claim = trace: evaluate agent claims against tool evidence. Zero deps.
Author-email: Jérémy Renoult <trinity@trinity-organism.com>
License: MIT
Keywords: agents,evaluation,harness,claim-trace,llm
Classifier: License :: OSI Approved :: MIT License
Classifier: Programming Language :: Python :: 3
Classifier: Topic :: Software Development :: Quality Assurance
Requires-Python: >=3.10
Description-Content-Type: text/markdown
License-File: LICENSE
Dynamic: license-file

# claim-trace

**claim = trace** — evaluate an agent (or human) claim against **measured evidence only**.

Zero dependencies. One idea: *PASS only when the tool trail supports the sentence.*

```bash
pip install claim-trace   # or: pip install -e .
claim-trace demo
```

## Why

Providers report models writing most production code. The scarce skill is not keystrokes — it is **directing, harnessing, and sealing** model labor. This tiny library is the evaluation primitive:

| Verdict | Meaning |
|---------|---------|
| **PASS** | every requirement has matching evidence |
| **FAIL** | missing key or mismatched value |
| **UNKNOWN** | claim has no requirements (ill-formed) |

Narration without evidence **cannot** PASS.

## 90-second demo

```bash
claim-trace demo
# → verdict PASS with evidence from gh/pytest *as data*, not as story
```

Custom:

```bash
claim-trace evaluate \
  --claim-json '{"text":"skill-lint is public","requires":[{"key":"private","equals":false}]}' \
  --evidence-json '[{"key":"private","value":false,"source":"gh api"}]'
```

## Not this package

- Not a full agent OS  
- Not your private harness, skills, or host config  
- Not a substitute for pair-screen on a production body  

It is the **public atom** of a larger private practice: claim = trace.

## License

MIT · [trinity-organism](https://github.com/trinity-organism)
