From natural language to enforceable mandate
Question answered: Do I need to learn a policy language? How does a user intent become a structured mandate?
You describe what you want in natural language. An agent turns it into a policy — typed constraints the policy engine can check — and the user’s authorized intent references that policy. Together they are the mandate.
A user might say: “I want a green sweater, size medium, 100% wool, max $100, from an approved merchant, delivered by Friday.” Each clause becomes one constraint:
| Clause | Constraint type |
|---|---|
| ”green” | attribute match |
| ”size medium” | attribute match |
| ”100% wool” | attribute match |
| ”max $100” | value bound |
| ”from an approved merchant” | counterparty restriction |
| ”delivered by Friday” | temporal condition |
The product taxonomy is what makes this work. It maps natural-language concepts to the specific, verifiable attributes the policy engine can check — “green” resolves to a color attribute, “100% wool” to a material attribute. Without the taxonomy, natural language is just text; with it, every clause maps to a typed field the policy engine can test against extracted evidence.
Compilation happens through the MCP. The loop is the one any compiler gives you: the agent drafts a policy, and the compiler either accepts it or returns a compile error. The agent fixes and retries until it compiles. An accepted policy is stored, ready for an intent to reference.
The user never sees the policy language. They see the natural-language description and the structured mandate it produced. The gap between the two is the taxonomy and the compiler.