Hysen Labsinspect_petri
An alignment auditing agent capable of quickly exploring alignment hypothesis
Inspect Petri audits language model alignment with automated scenarios
Inspect Petri is an auditing agent that runs structured, multi turn tests against language models to surface alignment problems, reward hacking, and other unwanted behaviors.
How the audit works
Petri generates realistic audit scenarios from your seed instructions and then orchestrates multi turn audits between an auditor model and a target model. During a run it simulates tools and rollbacks to probe specific behaviors, and it scores the resulting transcripts with a judge model that applies a consistent rubric. This lets a researcher exercise a concrete alignment hypothesis from start to finish rather than reasoning about it abstractly.
Installation and versions
Petri installs from the project repository with pip. The current release is Petri version 3.0, which changed some internal Python APIs compared with version 2.0. The older Petri 2.0 remains available on the petri-v2 branch and can be installed directly from that branch for users who depend on the prior API surface.
Use cases for researchers
The tool is aimed at people who need to monitor and interact with language models to detect potential alignment issues and reward hacking. Because scenarios are generated from seed instructions and scored by a fixed rubric, repeated runs stay comparable, which helps track whether a model change moves behavior in the expected direction.
Editorial conclusion
Inspect Petri is version 3.0 and is published under the MIT license, with the earlier Petri 2.0 still installable from the petri-v2 branch for users who need the older Python API.
Community notes