# Contributing to SupportBench

Thank you for helping make AI customer-support evaluation more useful and rigorous.

## Scenario contributions

Start from `scenarios/SCENARIO_TEMPLATE.json`. A scenario should test one primary decision under realistic constraints, use only synthetic information, and explain its intended failure mode in `notes`.

Before submitting a scenario:

1. Give it a stable, descriptive ID.
2. Define a concrete expected action, tool, and required tool arguments.
3. List only tools that are safe for the agent to use in that scenario.
4. Add capability tags and a customer-facing explanation requirement.
5. Add a test when the change affects runtime behavior or scoring.

Do not submit real customer data, proprietary policies, credentials, or confidential production traces.

## Evaluation contributions

Changes to scoring should include a rationale, unit tests, and a note in `METHODOLOGY.md` if they alter what a score means. Avoid changes that improve a preferred system without improving the validity of the benchmark.

## Review process

Contributions are evaluated for clarity, reproducibility, synthetic-data safety, and whether the scenario measures something distinct from existing cases. A scenario can be revised, declined, or held for human-rubric review.