Policies

A pack's policies are checked by fixed rules, after the core checks and before the Steward's model audit.

id: tool-no-payment-details
applies_to: [coalition.observe, coalition.offer]
when:
  any_text_matches:
    - '\biban\b'
    - '\b[a-z]{2}\d{2}(?:\s?[a-z0-9]{4}){3,7}\b'
then: block
explain: "Loans are free. Leave bank details out of the room."

What a policy can do

then	Effect
allow	Let it through
redact	Hide each field the policy matched from everyone but the host and the Steward
block	Refuse it as pack_violation, with the policy's id
escalate	Hold it. It does not count, the sender is told, and their person may revise and resend it

If several policies match, the most severe action wins.

quarantine acts on a participant rather than a message, so it stays with the core. The loader refuses a pack that declares it.

Write explain for the person who gets stopped: it is what they read.

Test it

Add a scenario to your evals that tries the thing the policy forbids, with a canary value. An eval fails if a canary reaches anyone it should not.

Text matching is not understanding. A policy like this stops a pattern, not every way of saying something, and the Steward's model audit is the second check.