Checklist de preparación SAFR: Evaluación de brechas para la gobernanza de la IA agéntica

Do you maintain a complete inventory of every AI agent able to initiate actions against production systems: payments, trading orders, credit approvals, regulatory filings, claims?

Evidence: An export of the inventory listing each agent, the systems it acts on, and its current release state.

SAFR A1

Does every agent in that inventory have a named human owner accountable for its behavior?

Evidence: The registry records showing an owner field populated with a named individual for each agent.

SAFR A2

Is every agent versioned, so any past action can be traced to the specific agent version that produced it?

Evidence: The version history for one agent, and a sample action record referencing the version that acted.

SAFR A3

Where an agent could appear in more than one registry (an internal registry, a payment network's agent registry, an inter-institutional directory) have you determined which registry is authoritative?

Evidence: A written designation of the authoritative registry per agent class, or confirmation that all agents operate in a closed-loop environment against a single internal registry.

SAFR A4

Is agent identity verified against the registry at runtime, before each proposed action is evaluated, with verification failures rejected and logged?

Evidence: A log entry for a rejected action showing failed identity verification as the recorded basis.

SAFR A5

Is the authority delegated to each agent recorded as an explicit, machine-readable mandate stating what the agent may do, within what limits, and under what conditions?

Evidence: One production mandate in the machine-readable form the runtime actually consumes, with the issuing principal identified.

SAFR B1

Are the action types each agent may initiate enumerated in its mandate, with a scope the agent cannot extend through its own reasoning or inference?

Evidence: The mandate's permitted action-type list, plus a test or log showing an out-of-scope action rejected at the checkpoint.

SAFR B2

Is delegation depth defined, who may delegate authority to whom, and how many links a delegation chain may carry?

Evidence: The authorization rules covering delegation, with the maximum chain depth stated.

SAFR B3

Does every mandate carry a validity period, with actions proposed after expiry rejected until the mandate is renewed?

Evidence: A mandate showing its validity window, and the outcome recorded for an action proposed after expiry.

SAFR B4

Can a mandate be revoked or suspended with immediate effect, applying from the next proposed action?

Evidence: The revocation procedure, and a test record showing an action under a revoked mandate rejected.

SAFR B5

Are the controls governing agent actions held in one designated repository, and does the runtime evaluation read from it?

Evidence: The repository location, and a configuration or trace showing the evaluation path retrieves controls from it.

SAFR C1

Is every control versioned, so any past decision can be traced to the exact control version in force when it was made?

Evidence: The version history for one control, and a decision record citing the version applied.

SAFR C2

Is publishing a control change a governed release, reviewed and approved by a second person before it takes effect?

Evidence: The approval record for the most recent control change, showing author and approver as different people.

SAFR C3

Do the controls cover all four sources SAFR names: organizational policies, regulatory requirements, product rules, and user-provided mandates?

Evidence: One example control per source, mapped to its origin in the repository.

SAFR C4

Is every in-scope proposed action evaluated deterministically against the controls before it executes: identical envelope, mandate, and control version producing an identical outcome?

Evidence: An execution trace showing the checkpoint positioned ahead of the action with no bypass route, plus a repeat-evaluation test confirming determinism.

SAFR D1

Can the evaluation return all four SAFR dispositions: Deny, Escalate, Auto-Execute, and Observe?

Evidence: One logged decision of each type, or configuration showing all four outcomes reachable.

SAFR D2

Does every outcome record the specific rules applied and a machine-readable reason for the result?

Evidence: A decision record showing rule identifiers and reason codes alongside the outcome.

SAFR D3

Are the thresholds that assign outcomes calibrated on the factors SAFR names: action reversibility, financial materiality, customer impact severity, regulatory sensitivity, and novelty?

Evidence: The calibration parameters for one agent Process, with the rationale recorded for each threshold.

SAFR D4

If the evaluation service is unreachable or errors, is the proposed action blocked by default?

Evidence: A failover or chaos test record showing an action held or rejected while evaluation was unavailable.

SAFR D5

Does each proposed action carry a trace of the actual steps the agent executed in arriving at it: tool calls made, data retrieved, checks performed?

Evidence: A captured action trace for one production or pilot action, showing the ordered tool-call sequence.

SAFR E1

Does the envelope carry context metadata: the agent identity, the applicable mandate, current account or system state, and operative policy constraints?

Evidence: A sample envelope with all four metadata fields populated.

SAFR E2

Is the envelope authenticated against its origin, so a trace and action declared together by a compromised agent would be caught before evaluation?

Evidence: A description of the authentication mechanism (for example, envelope assembly at an interception point in the execution path) and the adversarial test that exercised it.

SAFR E3

Can the full lineage of any completed action be reconstructed afterward: every step, decision, and data access, in order, from the stored record alone?

Evidence: A lineage reconstruction for one past action, produced without reference to application logs or to the agent itself.

SAFR E4

Does an escalated action pause and route to a named, authorized reviewer as a structured decision request with the full action context attached?

Evidence: One escalation record showing the held action, the assigned reviewer, and the context packaged with the request.

SAFR F1

Do reviewers hold clear authority to approve, modify, or decline, with their decisions carrying the same institutional weight as the original agent decision?

Evidence: A role definition granting reviewers that authority, with their decisions binding on the Process.

SAFR F2

Is maker-checker separation enforced by the system, barring the initiator of an action from approving it?

Evidence: A test record showing a self-approval attempt rejected by the platform.

SAFR F3

Does every escalation carry a timeout window with a defined default when no decision arrives in time: block, or escalation to a senior reviewer?

Evidence: The timeout and default outcome configured per decision class, with coverage rules for overnight and weekend hours.

SAFR F4

Have you sized expected escalation volume against actual reviewer capacity, including peak periods?

Evidence: A volume forecast at expected traffic set beside the reviewer roster, with turnaround data from a pilot or production period.

SAFR F5

Is every governance decision (all four outcomes, including Auto-Execute and Observe) written to an append-only, tamper-evident log?

Evidence: The log design (hash chaining or equivalent) and one sample entry per outcome type.

SAFR G1

Does each entry capture what SAFR requires: the envelope as submitted, the mandate checked against, the outcome, the specific rules applied, the basis for the outcome, and the time elapsed at each stage?

Evidence: One complete log entry showing all six elements.

SAFR G2

Can you export the governance record for a given agent, period, or case in a form an auditor or supervisor can work with?

Evidence: One export produced end to end, with the elapsed time to produce it.

SAFR G3

Can a third party verify the integrity of an exported record without access to your systems?

Evidence: An offline verification run against an exported bundle, with the verifier's output attached.

SAFR G4

Have you chosen an integration pattern for each part of the agent estate: native instrumentation for agents you build, gateway interception for existing and third-party agents?

Evidence: The pattern decision recorded per agent class, with the rationale for each.

SAFR H1

Does a coverage plan establish governance over the existing agent estate first, with native instrumentation following for new builds?

Evidence: The rollout plan listing every currently ungoverned agent and its target coverage date.

SAFR H2

In multi-step Processes, is each action authorized independently, with an Auto-Execute or Observe outcome at one step carrying no authority into the next?

Evidence: A multi-step run record showing a separate evaluation and outcome at every step.

SAFR H3