# falsepass

> Probe the reward code of RL environments and eval tasks (verifiers / Prime Intellect Environments Hub, Inspect AI, HUD) for false passes (a known-bad completion is rewarded) and false fails (a known-good one is not). CLI, MCP server, GitHub Action and an agent skill. Apache-2.0.

Install: `uvx falsepass demo` (PyPI). MCP: `uvx --from "falsepass[mcp]" falsepass mcp`. Output is a list of leads; reproduce one and read the reward code before calling it a defect.

## Docs

- [Guide](https://super-eric.github.io/falsepass/docs/index.md): Install, quick start, every command and check, and how to go from a probe run to a report.
- [Using falsepass from an agent](https://super-eric.github.io/falsepass/docs/agents/index.md): MCP server, agent skill, JSON output and exit codes: everything a coding agent needs to audit a grader with falsepass.
- [使用指南](https://super-eric.github.io/falsepass/zh/docs/index.md): 安装、快速上手、全部命令和检查项，以及从探测结果到报告的流程。
- [Output files](https://super-eric.github.io/falsepass/docs/output/index.md): The files an audit writes, the JSON formats and the exit codes.
- [Severity scale](https://super-eric.github.io/falsepass/docs/severity/index.md): How findings are rated: Critical, High, Medium, Low.
- [Audit skill](https://super-eric.github.io/falsepass/docs/agents/skill/index.md): A ready-made audit procedure for coding agents, in the Agent Skills format.

## Writing

- [Your reward function is an assertion. I tested a few dozen of them like one.](https://super-eric.github.io/falsepass/blog/reward-functions-are-assertions/index.md): Testing the graders of RL environments the way a test engineer tests an assertion: the defects that kept turning up, what a real model did with them, and what to check before you publish an environment.

## Optional

- [Source](https://github.com/Super-Eric/falsepass)
- [Changelog](https://github.com/Super-Eric/falsepass/blob/main/CHANGELOG.md)