Petri (Inspect Petri)
Project record
Maintained by Meridian Labs6248
Petri (Parallel Exploration Tool for Risky Interactions) is an open-source tool for automated alignment audits of language models. An auditor model drives multi-turn conversations with a target model from seed instructions, simulating tools and rolling back turns, and a judge model scores the transcripts. Anthropic released it in October 2025 and handed its development to Meridian Labs, an AI evaluation nonprofit, in May 2026 with the release of Petri 3.0. The repository moved from safety-research/petri to meridianlabs-ai/inspect_petri.98731
- Repository: Repository (external site: github.com)
- Documentation: Documentation (external site: meridianlabs-ai.github.io)
- License: LICENSE (MIT) (external site: github.com)
- Release notes: Original release post (Anthropic) (external site: anthropic.com)
- Release notes: Petri 3.0 announcement (Meridian Labs) (external site: meridianlabs.ai)
Availability and license
Overall availability
Source code is on GitHub and the package is on PyPI as inspect-petri. Running audits requires API access to the auditor, target, and judge models, for example Anthropic or OpenAI API keys.143
Availability is separate from permission: read the license before using or redistributing.
Component reuse rights
- weights
- Unknown — no complete fact-level rights review
- code
- Reviewed qualifying license recorded — check scope and conditions
- data
- Unknown — no complete fact-level rights review
- documentation
- Unknown — no complete fact-level rights review
No complete system-rights review is recorded for this release.
Public materials checklist
| Item | Status | Notes and evidence |
|---|---|---|
| CodeIs the evaluation code published? | Public | Published on GitHub under the MIT License and on PyPI.124 |
| Tasks / dataAre the tasks or test data available? | Public | The package ships more than 170 built-in seed instructions and 38 built-in judging dimensions; users can also supply their own.3 |
| MethodologyIs the method for scoring described? | Public | The documentation describes the auditor, target, and judge roles, the audit options, and the scoring: each dimension gets a 1-10 score with a written justification that cites specific messages.3 |
| ReproducibilityAre instructions for reproducing results published? | Public | The documentation gives install steps and inspect eval commands that select seeds and assign a model to each role. Results also depend on which auditor, target, and judge models are chosen.3 |
| LimitationsAre known limitations documented? | Partial | Anthropic's October 2025 release posts discuss limits of the original version: scenarios that may reveal they are tests, inconsistent judge scores, a small pilot set of scenarios, and metrics that cannot capture everything that matters. The current documentation adds a responsible-use note on provider policies but no updated limitations section was found.9103 |
What it is useful for
Run and use notes
- Petri is built on the Inspect AI evaluation framework and requires Python 3.12 or later. The latest PyPI release at review was 3.1.1; Petri 2.0 remains available on a separate branch.341
- The documentation warns that the full default audit (all built-in seeds, up to 30 turns each, scored by a frontier judge model) typically takes a few hours and incurs substantial API costs, and suggests starting with one tag or a small sample limit.3
Organization context
Provenance and derivatives
Petri was developed by Anthropic's alignment team and released in October 2025; Petri 2.0 followed on January 22, 2026. In May 2026 Anthropic handed development to Meridian Labs, which released Petri 3.0 with a restructured auditor/target design. Meridian Labs states that Anthropic will continue to support Petri.91187
U.S. eligibility
Eligible · basis: U.S.-governed project
Meridian Labs maintains Petri: Anthropic's May 2026 post says it handed development to Meridian Labs, the 2026 copyright in the MIT license is held by Meridian Labs, and Meridian Labs is the PyPI author. Meridian Labs describes itself as a 501(c)(3) nonprofit and gives a Boston, Massachusetts address. The documentation describes Petri as a collaboration between Meridian Labs and the UK AI Security Institute's Red Team; this catalog treats Meridian Labs, which holds the repository and publishes releases, as the maintaining entity. Anthropic originated the project. This catalog did not check Meridian Labs' tax status in an IRS listing.824563
Sources
This listing is not an endorsement, a safety assessment, or a federal approval.