ADR 0002 — Workload tiers and risk tiers

Status

Historical acceptance-design vocabulary; the runnable fixture is retired. The operator-reported architecture now describes a generic host, operator-core VM, logical realms grouping workloads, and individual agent sandboxes. Realm grouping is not an independently proven isolation boundary.

Context

Historically, proposed lab workloads varied in risk. Over-controlling fictional lab work wasted effort; under-controlling production or sensitive workloads would be dangerous.

Decision

Separate runtime tiers (where a workload runs) from risk tiers (what controls it must carry):

• Historical proposed runtime tiers: process lab, reconciled long-running service, and ephemeral microVM job with denied egress and verified teardown. These are not current implementation claims.
• Risk tiers: L0–L5, escalating controls by autonomy, data sensitivity, and blast radius.

Consequences

• Risk classification and runtime isolation are distinct concerns.
• Promotion gates require observed evidence, not fictional acceptance outcomes.
• No production or inter-realm proof follows from this historical tier taxonomy.