We talked about the verification of the environment.

We arrived at the two conclusions:

1. The environment is a program. It will be run many, many times while training the agent. Thus, it can be battle-tested with runtime assertions. Such assertions would crash the environment, and we would inspect the stack trace & the logs to understand the root cause of the bug.

2. The environment is our model of the real world. Thus, it encodes our assumptions. If only a small number of people are responsible for verifying the environment, it's going to conform to the assumptions of those people - but it might not be enough. Indeed, the more people we can involve in the development of the environment, the better. So it makes sense to open-source the basic building blocks of the environment (but not the complete environment), so that other people could use those building blocks in their apps, and report bugs / open pull requests with fixes.