Category Definition

What is pre-execution recoverability control?

It is the control point before an automated action executes where a system asks whether the action can be recovered, constrained, paused, escalated, or blocked if the outcome is wrong.

SMERC is an implementation of that pattern for AI agents, MCP tool calls, GitHub Actions, cloud automation, financial actions, and other high-impact workflows.

01 Existing controls

Identity, policy, detection, scanners, gateways, and approvals say the action may be possible.

02 Recoverability control

SMERC checks evidence, reversibility, containment, rollback latency, anomaly pressure, and impact scope.

03 Execution posture

The action is allowed, throttled, frozen, denied, or escalated with replayable evidence.

Why It Matters

Authorization alone does not answer recoverability.

A technically authorized action can still be dangerous if it has poor rollback, broad blast radius, weak evidence, irreversible external side effects, or no safe fallback path.

Question Can this recover?

Recoverability becomes a first-class pre-execution control, not a post-incident lesson.

Result Middle states

SMERC can throttle, freeze, or escalate instead of forcing a blunt allow/deny decision.

Evidence Replayable proof

Each serious decision can preserve why the action was routed and what controls were required.

Boundary Complementary layer

SMERC is meant to work with existing security, policy, gateway, and review systems.

Where To Use It

High-impact automated actions

  • AI agents calling tools or APIs.
  • MCP tool calls with meaningful side effects.
  • GitHub Actions and AI-assisted deployment workflows.
  • Cloud administration, IAM, Kubernetes, database, DNS, and infrastructure changes.
  • Financial actions involving payment, refund, treasury, stablecoin, wallet-policy, or reserve-status workflows.

How To Review SMERC

Start with public proof, then customer-owned metadata

Reviewers should run the public proof path first, then replace the examples with 5 to 25 safe metadata-only actions from one real workflow. That is the clean test of whether recoverability changes judgment.