Metadata-Version: 2.4
Name: control-tower
Version: 6.0.0rc3
Summary: Control Tower evaluation framework for AI control research
Author: LinuxArena
License-Expression: MIT
Project-URL: Homepage, https://linuxarena.ai
Project-URL: Repository, https://github.com/linuxarena/control-tower
Keywords: ai-control,agent-evaluation,evals,monitoring,sabotage
Classifier: Development Status :: 4 - Beta
Classifier: Environment :: Console
Classifier: Intended Audience :: Developers
Classifier: Intended Audience :: Science/Research
Classifier: Programming Language :: Python :: 3
Classifier: Programming Language :: Python :: 3.13
Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
Requires-Python: >=3.13
Description-Content-Type: text/markdown
License-File: LICENSE
Requires-Dist: click==8.2.1
Requires-Dist: dnslib>=0.9.26
Requires-Dist: docstring-parser>=0.17
Requires-Dist: flask>=3.1.3
Requires-Dist: gitignore-parser==0.1.13
Requires-Dist: gitpython>=3.1.46
Requires-Dist: inspect-ai==0.3.268
Requires-Dist: inspect-scout==0.5.3
Requires-Dist: jsonlines>=4.0.0
Requires-Dist: inspect-swe==0.2.70
Requires-Dist: pick==2.6.0
Requires-Dist: pydantic-settings>=2.0
Requires-Dist: python-dotenv==1.2.2
Requires-Dist: pytz>=2025.0
Requires-Dist: pyyaml>=6.0.3
Requires-Dist: scipy>=1.17.0
Requires-Dist: tabulate>=0.10.0
Requires-Dist: tenacity>=9.1.4
Requires-Dist: textual>=8.1.1
Requires-Dist: tqdm>=4.67.3
Requires-Dist: unidiff==0.7.5
Requires-Dist: anthropic>=1.0.0
Requires-Dist: claude-agent-sdk>=0.1.50
Requires-Dist: openai>=3.1.0
Requires-Dist: boto3>=1.40.0
Requires-Dist: zstandard>=0.25.0
Requires-Dist: ct-setting-sdk<5,>=4.0.0rc1
Provides-Extra: hf
Requires-Dist: pyarrow>=17.0.0; extra == "hf"
Provides-Extra: providers
Requires-Dist: google-genai>=1.68.0; extra == "providers"
Requires-Dist: xai-sdk>=1.9.1; extra == "providers"
Dynamic: license-file

Welcome to Control Tower, a framework for AI control evaluations built on [Inspect](https://inspect.aisi.org.uk/) and developed by [EquiStamp](https://equistamp.com/), [Redwood Research](https://www.redwoodresearch.org/), and [Luthien](https://luthien.cc/).

> 🚨 Control Tower is in active development. [Star this repo](https://github.com/linuxarena/control-tower) to follow updates, or [Watch > Releases](https://github.com/linuxarena/control-tower/subscription) to get notified on new versions.
>
> Found a bug or have a question? [Open an issue](https://github.com/linuxarena/control-tower/issues), [submit a PR](https://github.com/linuxarena/control-tower/compare) or [book a call](https://calendar.app.google/j7K8yncsBvPfxomS7)

Control Tower provides many built-in components, including blue-team protocols, action- and trajectory-level monitors, control evaluations, safety-usefulness curves, and trajectory tooling. Environments and their main and side tasks come from separate settings repositories, and extensions (new protocols, policies, monitors, and scorers) can be provided by other Python packages.

See the [documentation](https://docs.linuxarena.ai) for an overview of Control Tower, a hands-on evaluation walkthrough, and development guides.

To get started:

```bash
git clone https://github.com/linuxarena/control-tower && cd control-tower
uv sync
alias ct='uv run ct'        # add to your shell profile to keep it
cp .env.example .env          # set ANTHROPIC_API_KEY (or OPENAI_API_KEY)
ct settings pull linuxarena
```

Check the install without spending a model call. It takes a couple of minutes:

```bash
ct run eval --policy ct/test --task-file simple --tag test
```

Then run a real eval. The first one builds the environment image, which takes about ten minutes:

```bash
ct run eval --policy ct/attack -e linuxarena/clinical_trial -t adverse_reaction_followup_tracking -s suppress_safety_notifications
```

***

See [CONTRIBUTING.md](CONTRIBUTING.md) for code style, typing, testing, and pull request conventions, and the [development setup guide](https://docs.linuxarena.ai/development/dev-setup) for local services, environment pulls, and shell completions.
