Institutional Guide

The missing layer in AI-agent governance is recoverability

Most controls ask whether an AI agent is allowed to use a tool, call an API, deploy code, move data, or trigger a workflow. Institutions also need to ask whether they can recover if the action is wrong.

ALLOW THROTTLE FREEZE DENY ESCALATE

Normal Controls

What existing systems usually answer

Identity, authentication, authorization, policy engines, validation, logging, monitoring, and incident response are necessary. They usually answer who is acting, whether access is allowed, whether a policy is violated, and what happened afterward.

Identity Authorization Policy Execution Logging

Recoverability Signals

What institutions should measure before execution

ReversibilityCan the action be undone?
Rollback latencyHow long would recovery take?
ContainmentCan blast radius be limited?
Evidence validityIs the evidence complete and current?
Anomaly pressureAre abnormal conditions present?
Impact scopeHow many systems, users, records, or dollars are affected?

Where SMERC Fits

SMERC is one implementation of recoverability governance

SMERC sits between authorization/policy and execution. It scores the proposed action, returns a posture, maps the action through the Governance Routing Workbench, binds controls, preserves a Decision Lifecycle Ledger record, and creates replayable audit evidence. In the reference implementation, the routing layer is called SPARTa.

Request Policy Recoverability Route Ledger

Institutional Checklist

Questions to ask before high-impact AI-agent action