In-memory engine

1. Demos and tests that need a governed connection in one command, with no server to install.
2. The worked example a new connector is written against.
3. A home for the certification corpus: 124 cases, strict, including around eighty adversarial ones.

​

Configuring one

connections:
  - name: staging
    engine: memory
    policy: ./policies/staging.yaml
    seed:
      resources:
        - kind: collection
          path: [shop, orders]
          fields:
            - { name: id, type: string, identifying: true }
            - { name: tenant_id, type: string }
            - { name: status, type: enum, enum_values: [paid, refunded, cancelled] }
          rows:
            - { id: ord_1, tenant_id: acme, status: paid }
            - { id: ord_2, tenant_id: globex, status: paid }

​

The command grammar

{
  "verb": "update",
  "resource": "orders",
  "filter": { "id": "$0" },
  "set": { "status": "refunded" }
}

• A key the verb does not use — a find carrying set — is refused, not ignored. Ignoring it is how a write gets reviewed as a harmless read.
• A filter on a field the resource does not have is refused, because it would otherwise read as a precise lookup while matching nothing.
• A malformed parameter token ($x, $-1) is refused rather than treated as a literal string.
• An operator it cannot normalise becomes an unknown predicate carrying a hazard, which makes the operation unbounded — the honest representation of “well-formed, and I cannot reason about it”.

​

What it certifies