Use a realistic task with sample material
- Use copies or made-up sample data.
- Keep network and messaging access off unless the test requires it.
- Ask the Bot to describe the files, tools, and approvals it expects to use.
- Verify the claimed output exists and can be opened.
- Check logs and the destination directory for unrequested changes.
- Record the Hermes version, operating system, exact project version, test date, and result.
Record failed tests
Stop when setup is confusing, a tool is missing, a credential request is surprising, or the result differs from the README. Record the problem and resolve it before adding more access.
Add one type of access at a time
Add one type of access at a time. Repeat the test when the source, AI model, tools, outside connections, or schedule changes.