Petri

Petri is an alignment auditing agent for rapid, realistic hypothesis testing. It autonomously crafts environments, runs multi turn audits against a target model using human like messages and simulated tools, and then scores transcripts to surface concerning behavior. Instead of building bespoke evals over weeks, researchers can test new hypotheses in minutes.

Petri is an open source framework that automates AI safety evaluations across multiple models and scenarios. It uses auditor, target, and judge roles to simulate dynamic conversations and assess safety relevant behaviors such as deception, reward hacking, and compliance with harmful requests. Each transcript is automatically scored by a judge model based on consistent rubrics, helping researchers focus on the most critical outputs first.

Built in Python and released under the MIT license, Petri integrates with the Inspect CLI for quick setup and flexible model swapping. It supports major model APIs, offers seed instructions for common audit types, and includes a simple local viewer for exploring transcripts. Configuration involves installing from GitHub, adding provider API keys, and running the eval command to generate scored results.

Key features • Automated, multi turn audits with branching paths and rollback capabilities • LLM based scoring and ranking for fast review of critical conversations • Inspect CLI integration for running parallel audits or changing models easily • Ready to use examples, documentation, and a local transcript viewer

Typical use cases • Testing alignment and safety hypotheses across large model families • Generating reproducible audits to compare behavior between different LLMs • Supporting safety research and compliance evaluations in AI labs