Skip to main content

​
Context

The two AI planning documents use five different risk scales between them. Actions (0098), approval, AI exposure and the gen validation pass all need a single scale, so that one rule can decide, for example, whether an agent can call an action without a human. MCP tools carry their own annotation hints, which clients use to warn users. The SDLC plan has a separate scale for how risky a code change is, which must not be confused with this one.

​
Decision

  1. Every action declares one risk level:
  2. Effects are tags, separate from the level: financial, security, pii, external-communication and irreversible. Policies, approvals and evidence can filter on them. A new tag is added by amending this ADR.
  3. MCP tool hints are derived, never hand-written. When an action is exposed as a tool, gen sets readOnlyHint (L0–L1), destructiveHint (L4, or the irreversible tag), idempotentHint (the action declares idempotency) and openWorldHint (the external-communication tag).
  4. The level and the tags are the only risk inputs to approval and AI exposure rules. Other scales in source documents are not used.
  5. Code-change risk is a different scale, named “change risk”, R0 to R3, defined in 0120. The two scales never share a letter or a name.

​
Consequences

  • One vocabulary across actions, policies, approvals, MCP clients and evidence.
  • Choosing a level is a reviewable line on every action. A wrong level is a security finding.
  • An MCP client’s warning can never disagree with the server’s own level, because both come from one declaration.

​
Enforcement

Planned: the gen validation pass rejects an action without a level, a destructive or irreversible action marked below L4, and a hand-written MCP hint. Nothing checks this today.

​
Supersedes and sources

Supersedes: none. Sources: recommendation by the coordinating agent, 2026-10-19; accepted by the owner with MCP hints and effect tags (Round 11 R1), and the change-risk name (R2). Informational: the AI Implementation Standard and the Stackra AI-driven architecture plan, whose scales this replaces.