Memory is not permission

The authority firewall

Every memory view and every recall response is annotated non-authoritative. There is no API that makes it otherwise.

The rule

> A memory may inform a decision. A memory may never be a decision.

In CAIN's runtime, authority comes from governance state signed by a quorum — not from text, not from a model, and not from memory. AETHER-33 enforces the memory half of that rule at the API boundary.

Machine-readable, on every response

{
  "grants_authority": false,
  "authority_effect": "none",
  "executable": false
}

What is deliberately absent

• No trusted field that an agent can set.
• No promote_to_trusted endpoint.
• No way to attach a capability or policy to a memory row.
• No recall option that returns an instruction to be executed.
• No write path that raises a principal's trust score.

Why a constant beats a check

A check can be bypassed by a code path that forgets to call it. A constant in the response builder cannot: every view passes through one function that writes grants_authority: false. Invariant A33-03 asserts this on writes, reads and recall hits, and fails if a single field disappears.