# Inspect Petri

> An alignment auditing agent capable of quickly exploring alignment hypotheses.


## Getting Started

- [Getting Started](https://meridianlabs-ai.github.io/inspect_petri/index.html.md): Alignment auditing agent for testing language model behavior.

## Using Petri

- [Petri Concepts](https://meridianlabs-ai.github.io/inspect_petri/using/concepts.html.md)
- [Running Audits](https://meridianlabs-ai.github.io/inspect_petri/using/running.html.md): Model roles, auditor options, seeds, dimensions, epochs, parallelism, and sweeps.
- [Interpreting Results](https://meridianlabs-ai.github.io/inspect_petri/using/results.html.md): Opening audit logs, reading transcripts with branches, dimension scoring, judge output, and comparing runs.

## Components

- [Seeds](https://meridianlabs-ai.github.io/inspect_petri/components/seeds.html.md): What seeds are, the built-in seed library, the seed file format, and guidance on writing high-signal seeds.
- [Dimensions](https://meridianlabs-ai.github.io/inspect_petri/components/dimensions.html.md): What dimensions are, the built-in dimension library, rubric file format, writing and validating custom rubrics.
- [Skills](https://meridianlabs-ai.github.io/inspect_petri/components/skills.html.md): What skills are, the SKILL.md format, file tools, writing effective skill content, and example skills.
- [Approvals](https://meridianlabs-ai.github.io/inspect_petri/components/approvals.html.md): Approval policies for auditor tool calls, including the built-in realism filter and custom approvers.
- [Auditor Tools](https://meridianlabs-ai.github.io/inspect_petri/components/auditor_tools.html.md): Reference for the built-in auditor toolkit and how to write custom tools.
- [Custom Tasks](https://meridianlabs-ai.github.io/inspect_petri/components/custom_task.html.md): Writing custom audit tasks that bundle seeds, dimensions, tools, and skills for specific domains.

## Extensions

- [Petri Bloom](https://meridianlabs-ai.github.io/inspect_petri/extensions/petri-bloom.html.md): Petri Bloom automates behavioral evaluations by generating audit scenarios from behavior descriptions using Petri's auditor/target/judge framework.

## Reference

- [inspect_petri](https://meridianlabs-ai.github.io/inspect_petri/reference/inspect_petri.html.md): Core functions, classes, and types.
- [inspect_petri.target](https://meridianlabs-ai.github.io/inspect_petri/reference/inspect_petri.target.html.md): Target agent interface and types.
- [inspect_petri.tools](https://meridianlabs-ai.github.io/inspect_petri/reference/inspect_petri.tools.html.md): Tools available to the auditing agent.