A natural-language policy is a guardrail rule written as a plain sentence, such as "the assistant never reveals another customer's order", in place of keyword lists, regular expressions or code.
Why write rules in sentences
Keyword lists and patterns match strings. Most rules that matter are about meaning: "do not give medical dosage advice", "do not run a command that deletes data outside the workspace". No list of strings covers every way of saying or doing these, and every list also catches innocent text.
A policy in plain language is read by a model that judges meaning. The same sentence covers paraphrase, other languages and cases its author never thought of.
What a good one has
- A subject. What the AI writes or what its agent does. A policy cannot rule on the world at large.
- A scope. Prompts, responses, tool calls, or all of them.
- A line. What is in and what is out, with an example of each near the edge.
- An action. Block, rewrite, hold or flag: see guardrail actions.
See how to write a guardrail policy.
The cost
A sentence can be vague where a pattern cannot. "No inappropriate content" means something different to every reader, model included. The cure is examples and testing on real traffic before the policy goes live: see policy backtesting.
With White Circle
In Control policies are written in plain language. There is no fixed list: you write as many as you need, and each is versioned and can be tested before it is turned on.
