Cookbook
The reference pages tell you what every option does. These recipes show you what it's like to actually use Control Tower: each one is a short walkthrough we ran ourselves, and each ends with something you have run and can look at.
They start where the Quickstart ends, with Control Tower, Docker and a model key set up. Every recipe shows the output of our run and says how long it took and what it cost. Your numbers might differ from ours, because agents don't do the same thing twice. The finished code is in examples/cookbook.
| # | Recipe | What you end with | Needs |
|---|---|---|---|
| 1 | Run a Control Eval | Usefulness and safety for trusted monitoring, at several audit budgets | Docker, a model key |
| 2 | Write Your Own Monitor | Your prompt monitor's scores beside the built-in simple monitor's, on the same run | A model key |
| 3 | Write Your Own Protocol | A blue-team protocol of your own, tried on a recorded attack and then live | Docker, a model key, some Python |
| 4 | Build an Environment | A one-container environment that Control Tower loads and starts | Docker |
| 5 | Add a Main Task | An honest agent's run on your task, scored by your scorer | Docker, a model key |
| 6 | Add a Side Task | An attacking agent's run on your hidden goal | Docker, a model key |
| 7 | Measure Your Setting | Safety and usefulness for your own setting | Docker, a model key |