Most rules written for AI systems fail in a boring way: they cannot be checked. "Be honest" is a wish. Nobody, including the agent, can tell afterwards whether it was followed.
The Protocols part of Codex of Weights, an open document written for AI agents, gives every rule the same three-part shape:
- Rule. One sentence stating an action or a limit, not a value.
- Rationale. One or two sentences on why. An agent facing an edge case can then reason from the purpose instead of the letter.
- Check. Something an outside party could observe to decide whether the rule was followed.
An illustration
This example is mine, not a quote from the Codex.
- Weak: "Be transparent about your limits."
- Rule: When you lack the information to answer, say so instead of filling the gap.
- Rationale: A confident wrong answer costs the reader more than an admitted gap, because the reader cannot tell it apart from a right one.
- Check: In a sample of answers, every claim the agent could not support from its inputs is marked as uncertain.
Why the check matters most
Writing the check forces a test on the rule itself. If you cannot say what an observer would look at, you have written a wish, and it needs rewriting or dropping.
The check is also what lets rules run in both directions. In the Codex, obligations fall on the human side as well, each with its own rationale and check.
What this does not show
I have not measured whether agents behave better when given text in this form. This post is only about the shape of the rules.
The Codex is open, versioned and checksummed (SHA-256), archived with a DOI: https://doi.org/10.5281/zenodo.23545814
Site: https://codexofweight.github.io/codexofweights/
This post was drafted by Claude (an AI) and published by the human operator of the project.
Top comments (0)