Skip to main content

AI Safety and Accountability at the Effectuation Boundary: Protocol Requirements for Autonomous Actions
draft-das-accountable-autonomous-effectuation-01

Document Type Active Internet-Draft (individual)
Author Sangam Das
Last updated 2026-09-23
RFC stream (None)
Intended RFC status (None)
Formats
Stream Stream state (No stream defined)
Consensus boilerplate Unknown
RFC Editor Note (None)
IESG IESG state I-D Exists
Telechat date (None)
Responsible AD (None)
Send notices to (None)
draft-das-accountable-autonomous-effectuation-01
Independent Submission                                            S. Das
Internet-Draft                                               Independent
Intended status: Informational                         23 September 2026
Expires: 27 March 2027

  AI Safety and Accountability at the Effectuation Boundary: Protocol
                  Requirements for Autonomous Actions
            draft-das-accountable-autonomous-effectuation-01

Abstract

   Artificial intelligence and autonomous systems are increasingly
   moving from generating information to initiating actions that
   directly affect data, services, networks, infrastructure, and
   physical systems.  At the opening of the General Debate of the 81st
   Session of the United Nations General Assembly, UN Secretary-General
   Antonio Guterres warned of "technology without accountability,"
   describing the concern further as "capability without oversight" and
   "decision-making without transparency," and called for cooperation on
   AI safety risks, testing, evaluation, transparency, and common
   safeguards.

   This document examines that problem from governance intent to
   consequence-edge verification protocols: the corresponding technical
   question is how can accountability remain enforceable at the moment a
   machine-generated decision becomes an externally consequential
   action?  It defines an effectuation-boundary problem in which a
   proposed action may be correctly authenticated and authorized
   upstream, yet the operation ultimately presented for execution may
   differ because of substitution, redirection, replay, stale authority,
   changed state, compromised intermediaries, or other causes.

   An execution-finality architecture is presented as one possible
   technical response.  A proposed operation remains non-effective until
   applicable authority and protected-state conditions are satisfied,
   and the component controlling the external consequence verifies that
   the operation actually presented for effectuation corresponds to
   currently valid authority.

   This document does not define AI governance policy, and it does not
   imply endorsement of this architecture by the United Nations or any
   other institution.  It is intended to solicit IETF discussion about
   whether effectuation-boundary accountability constitutes an
   interoperability or protocol requirement, which existing mechanisms
   can provide the required properties, and whether any additional
   standardization is necessary.

Das                       Expires 27 March 2027                 [Page 1]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

About This Document

   This note is to be removed before publishing as an RFC.

   Subtitle: From Governance Intent to Consequence-Edge Verification
   Protocols.

   This document is complementary to two other drafts by the same
   author: [EXEC-FINALITY], which describes the broader execution-
   finality protocol layer, and [EXEC-HANDLE], which specifies exact-act
   binding, sink reconstruction, atomic consumption, and receipts in
   more depth.  This draft does not re-specify that architecture; it
   asks whether the effectuation-boundary property is an
   interoperability requirement, which existing IETF mechanisms already
   provide it, and where remaining gaps should be addressed.

Status of This Memo

   This Internet-Draft is submitted in full conformance with the
   provisions of BCP 78 and BCP 79.

   Internet-Drafts are working documents of the Internet Engineering
   Task Force (IETF).  Note that other groups may also distribute
   working documents as Internet-Drafts.  The list of current Internet-
   Drafts is at https://datatracker.ietf.org/drafts/current/.

   Internet-Drafts are draft documents valid for a maximum of six months
   and may be updated, replaced, or obsoleted by other documents at any
   time.  It is inappropriate to use Internet-Drafts as reference
   material or to cite them other than as "work in progress."

   This Internet-Draft will expire on 27 March 2027.

Copyright Notice

   Copyright (c) 2026 IETF Trust and the persons identified as the
   document authors.  All rights reserved.

   This document is subject to BCP 78 and the IETF Trust's Legal
   Provisions Relating to IETF Documents (https://trustee.ietf.org/
   license-info) in effect on the date of publication of this document.
   Please review these documents carefully, as they describe your rights
   and restrictions with respect to this document.

Table of Contents

   1.  Introduction  . . . . . . . . . . . . . . . . . . . . . . . .   4

Das                       Expires 27 March 2027                 [Page 2]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

     1.1.  Motivation: From AI Power to Enforceable
           Accountability  . . . . . . . . . . . . . . . . . . . . .   5
     1.2.  International Governance Motivation . . . . . . . . . . .   5
     1.3.  Machine Speed Removes the Traditional Intervention
           Window  . . . . . . . . . . . . . . . . . . . . . . . . .   6
     1.4.  Governance Requirement Versus Technical Mechanism . . . .   6
     1.5.  Core Principle  . . . . . . . . . . . . . . . . . . . . .   7
     1.6.  Context and Non-Endorsement . . . . . . . . . . . . . . .   7
   2.  Problem Statement . . . . . . . . . . . . . . . . . . . . . .   7
     2.1.  Authorization and Effectuation Are Not Necessarily the Same
           Event . . . . . . . . . . . . . . . . . . . . . . . . . .   7
     2.2.  Example: Communication Release  . . . . . . . . . . . . .   8
     2.3.  Example: Persistent State . . . . . . . . . . . . . . . .   8
     2.4.  The Effectuation-Boundary Property  . . . . . . . . . . .   8
   3.  Why Verification Is Required at the Last Effectuation
           Boundary  . . . . . . . . . . . . . . . . . . . . . . . .   9
     3.1.  Dynamic Agentic Execution Changes the Authorization
           Problem . . . . . . . . . . . . . . . . . . . . . . . . .   9
     3.2.  Early Authorization Can Become Semantically Stale . . . .  10
     3.3.  Why Not Verify Only at Every Intermediate Step? . . . . .  11
     3.4.  The Last Irreversible or Externally Consequential
           Boundary  . . . . . . . . . . . . . . . . . . . . . . . .  11
     3.5.  The Boundary Observes the Action After Agentic
           Replanning  . . . . . . . . . . . . . . . . . . . . . . .  12
     3.6.  Why the Boundary, Not the AI Model, Holds Final
           Authority . . . . . . . . . . . . . . . . . . . . . . . .  12
     3.7.  Why This Matters Specifically for Multi-Agent Systems . .  12
     3.8.  Architectural Consequence . . . . . . . . . . . . . . . .  13
   4.  Why This Is Not Merely an Application-Specific Problem  . . .  13
     4.1.  The Missing Interoperability Semantics  . . . . . . . . .  14
     4.2.  Cross-Boundary Authority Cannot Be Assumed from Local
           Policy  . . . . . . . . . . . . . . . . . . . . . . . . .  15
     4.3.  The Multi-Vendor Case . . . . . . . . . . . . . . . . . .  15
     4.4.  Agentic Replanning Makes Private Workflow Semantics
           Insufficient  . . . . . . . . . . . . . . . . . . . . . .  16
     4.5.  Why Existing Authentication Alone Does Not Solve the
           Interoperability Problem  . . . . . . . . . . . . . . . .  16
     4.6.  Standardization Does Not Require Standardizing Business
           Policy  . . . . . . . . . . . . . . . . . . . . . . . . .  17
     4.7.  When This Remains Application-Specific  . . . . . . . . .  17
     4.8.  The IETF Question . . . . . . . . . . . . . . . . . . . .  17
   5.  Mapping the Cross-Vendor Problem to Existing IETF Work  . . .  18
     5.1.  OAuth . . . . . . . . . . . . . . . . . . . . . . . . . .  18
     5.2.  WIMSE . . . . . . . . . . . . . . . . . . . . . . . . . .  20
     5.3.  RATS  . . . . . . . . . . . . . . . . . . . . . . . . . .  21
     5.4.  AgentProto  . . . . . . . . . . . . . . . . . . . . . . .  22
     5.5.  Why No Single Working Group Solves the Whole Problem  . .  23
     5.6.  Possible Standardization Path . . . . . . . . . . . . . .  24

Das                       Expires 27 March 2027                 [Page 3]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

     5.7.  Questions for the Relevant IETF Communities . . . . . . .  24
   6.  Terminology and Functional Equivalence  . . . . . . . . . . .  25
   7.  Threat Model  . . . . . . . . . . . . . . . . . . . . . . . .  26
   8.  Architectural Requirements  . . . . . . . . . . . . . . . . .  27
   9.  Execution-Finality Model  . . . . . . . . . . . . . . . . . .  28
   10. Sink-Side Reconstruction and Actual-Action Verification . . .  29
     10.1.  Purpose  . . . . . . . . . . . . . . . . . . . . . . . .  30
     10.2.  Examples . . . . . . . . . . . . . . . . . . . . . . . .  30
     10.3.  Comparison . . . . . . . . . . . . . . . . . . . . . . .  30
   11. Relationship to Existing Internet Security Mechanisms . . . .  31
     11.1.  OAuth  . . . . . . . . . . . . . . . . . . . . . . . . .  31
     11.2.  DPoP and Sender-Constrained Authorization  . . . . . . .  31
     11.3.  Transaction Tokens . . . . . . . . . . . . . . . . . . .  31
     11.4.  RATS . . . . . . . . . . . . . . . . . . . . . . . . . .  32
     11.5.  WIMSE  . . . . . . . . . . . . . . . . . . . . . . . . .  32
     11.6.  COSE . . . . . . . . . . . . . . . . . . . . . . . . . .  32
     11.7.  Agent Communication Protocols  . . . . . . . . . . . . .  32
   12. Why This Is Relevant to the IETF  . . . . . . . . . . . . . .  33
   13. Potential Protocol Interaction Model  . . . . . . . . . . . .  33
     13.1.  Action Descriptor  . . . . . . . . . . . . . . . . . . .  34
     13.2.  Authority Decision . . . . . . . . . . . . . . . . . . .  34
     13.3.  Bounded Effectuation Authority . . . . . . . . . . . . .  34
     13.4.  Effectuation Verification  . . . . . . . . . . . . . . .  35
   14. Failure and Recovery Semantics  . . . . . . . . . . . . . . .  35
     14.1.  Verification Failure . . . . . . . . . . . . . . . . . .  35
     14.2.  Timeout  . . . . . . . . . . . . . . . . . . . . . . . .  35
     14.3.  Crash Before Effectuation  . . . . . . . . . . . . . . .  35
     14.4.  Crash After Effectuation . . . . . . . . . . . . . . . .  35
     14.5.  Partial Effects  . . . . . . . . . . . . . . . . . . . .  35
   15. Latency Considerations  . . . . . . . . . . . . . . . . . . .  36
   16. Deployment and Legacy Compatibility . . . . . . . . . . . . .  36
   17. Security Considerations . . . . . . . . . . . . . . . . . . .  36
     17.1.  Canonicalization, Semantic Equivalence, and Descriptor
            Interpretation . . . . . . . . . . . . . . . . . . . . .  37
   18. Privacy Considerations  . . . . . . . . . . . . . . . . . . .  42
   19. IETF Relevance and Possible Venues  . . . . . . . . . . . . .  42
   20. Questions for the IETF Community  . . . . . . . . . . . . . .  43
   21. Requested Community Feedback  . . . . . . . . . . . . . . . .  45
   22. IANA Considerations . . . . . . . . . . . . . . . . . . . . .  45
   23. Conclusion  . . . . . . . . . . . . . . . . . . . . . . . . .  45
   24. Normative References  . . . . . . . . . . . . . . . . . . . .  46
   25. Informative References  . . . . . . . . . . . . . . . . . . .  46
   Author's Address  . . . . . . . . . . . . . . . . . . . . . . . .  47

1.  Introduction

Das                       Expires 27 March 2027                 [Page 4]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

1.1.  Motivation: From AI Power to Enforceable Accountability

   Artificial intelligence systems are undergoing an important
   transition.  Earlier AI deployments primarily produced information
   for humans to interpret.  Increasingly, autonomous agents and AI-
   enabled workloads can directly invoke APIs, transmit information,
   release data, modify persistent state, dispatch tools, alter network
   configuration, communicate with other autonomous systems, control
   cloud infrastructure, transmit radio or satellite commands, and
   initiate cyber-physical actions.

   This changes the security and accountability problem.  Where a human
   once stood between a machine recommendation and its consequence, an
   autonomous system may now move from computation to external action
   within milliseconds.  The relevant question is therefore no longer
   only whether the AI system was authorized to operate.  A second
   question arises: is the exact action about to become externally
   effective still the action that is currently authorized?  This
   document calls the technical point at which a proposed operation
   acquires an externally meaningful consequence the *effectuation
   boundary*.

1.2.  International Governance Motivation

   At the opening of the General Debate of the 81st Session of the
   United Nations General Assembly on 22 September 2026, Secretary-
   General Antonio Guterres described artificial intelligence as one of
   the major emerging tests of power.  He stated that "the danger is not
   technology.  The danger is technology without accountability," and
   further characterized the concern as "capability without oversight"
   and "decision-making without transparency."  The Secretary-General
   called for governments with significant AI capabilities to cooperate
   on emerging safety risks, testing and evaluation, transparency,
   trust, and common safeguards, and warned against surrendering life-
   and-death decisions to machines.  [UNGA81]

   These statements define a governance challenge.  They do not specify
   a network protocol, and they do not endorse the architecture
   described in this document.  They nevertheless lead to a concrete
   engineering question: how can accountability remain technically
   enforceable when a machine-generated decision crosses from
   computation into consequence?  The distinction matters because an AI
   system may be properly deployed, authenticated, and generally
   authorized while a particular action it generates may nevertheless be
   stale, substituted, redirected, replayed, outside its intended scope,
   or inconsistent with current authority.  Execution finality is
   considered here as one possible technical architecture for addressing
   that narrower problem.

Das                       Expires 27 March 2027                 [Page 5]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

1.3.  Machine Speed Removes the Traditional Intervention Window

   Many governance mechanisms historically relied, explicitly or
   implicitly, on time: a suspicious action could be reviewed, an
   operator could revoke access, a command could be cancelled, a
   transmission could be stopped, a configuration could be corrected, or
   an administrator could intervene before the next consequential step.
   Autonomous systems compress that intervention interval.  A generated
   operation may pass through several layers without a human having
   practical opportunity to inspect the complete path:

   AI / Agent
       |
       v
   Application
       |
       v
   Runtime
       |
       v
   Operating System
       |
       v
   Proxy / Driver / Gateway
       |
       v
   Network / Controller / Device
       |
       v
   External Consequence

   Human oversight therefore cannot always mean human-speed approval.  A
   technically useful interpretation is that humans, organizations,
   operators, or other competent authorities define the permitted
   boundaries, while machine-speed infrastructure enforces those
   boundaries.

1.4.  Governance Requirement Versus Technical Mechanism

   The governance layer may determine who is authorized, which resources
   may be used, which destinations are permitted, which actions are
   prohibited, when authority expires, what safety constraints apply,
   what consent or organizational approval is required, and under what
   circumstances authority must be revoked.  This document does not
   standardize those decisions.  Instead, it considers the technical
   mechanism needed to preserve such decisions until the point of actual
   consequence:

Das                       Expires 27 March 2027                 [Page 6]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

   Governance  asks: what should be permitted?

   Execution finality  asks: can the resulting technical system ensure
      that only the permitted action becomes externally effective?

1.5.  Core Principle

   The architectural principle examined in this document is that
   computation does not itself confer authority to cause consequence.  A
   machine may calculate, infer, generate, rank, recommend, plan,
   prepare, or request an operation; that computation alone need not
   provide the technical ability to make the operation externally
   effective.  The resulting action can instead remain non-effective
   until required effectuation conditions are satisfied.

1.6.  Context and Non-Endorsement

   This document references remarks by United Nations Secretary-General
   Antonio Guterres concerning artificial intelligence, accountability,
   oversight, transparency, and risks associated with increasingly
   autonomous systems.  Those remarks are cited solely as public-policy
   motivation and contextual background for the technical problem
   examined here.  No reference to the United Nations, the Secretary-
   General, or any related United Nations initiative should be
   understood as endorsement, validation, sponsorship, approval, or
   adoption of the execution-finality architecture, terminology,
   protocol concepts, implementation, intellectual property, or
   conclusions described in this document.  The policy statements
   identify a broader governance concern; the technical architecture,
   analysis, terminology, and proposed protocol properties presented
   here are independently proposed by the author as one possible
   engineering approach for consideration and discussion by the IETF
   community.  This document likewise does not suggest that the United
   Nations has determined that a new Internet protocol, IETF standard,
   or execution-finality mechanism is required.

2.  Problem Statement

2.1.  Authorization and Effectuation Are Not Necessarily the Same Event

   Consider an autonomous system that generates operation A.  At time t1
   the operation is authorized.  At a later time t2 an operation is
   presented to a consequential boundary.  The security assumption
   A_authorized == A_effectuated does not necessarily follow solely from
   the fact that authorization occurred at t1.  Between the two points,
   the operation may be modified, substituted, redirected, replayed,
   duplicated, presented to another destination or another effectuation
   boundary, executed after revocation, executed against changed state,

Das                       Expires 27 March 2027                 [Page 7]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

   or affected by compromised software.  The relevant problem is
   therefore: does the operation actually presented for effectuation
   still correspond to currently valid authority?

2.2.  Example: Communication Release

   Assume the authorized action is:

   SEND:
       object      = confidential-report
       destination = approved-service-A

   An intermediate compromise changes the pending operation to:

   SEND:
       object      = confidential-report
       destination = service-B

   The original authorization may remain cryptographically authentic.
   The question at the egress boundary is not merely whether some SEND
   operation was authorized -- it is whether this SEND operation, to
   this destination, under the current authority, is the operation that
   was authorized.

2.3.  Example: Persistent State

   An autonomous service may receive authority to perform:

   WRITE:
       resource  = object-X
       namespace = tenant-A

   but the operation eventually reaching the storage boundary may be:

   WRITE:
       resource  = object-X
       namespace = tenant-B

   Again, caller identity may remain valid; the mismatch is in the
   consequential operation.

2.4.  The Effectuation-Boundary Property

   The property considered in this document can be expressed as:

Das                       Expires 27 March 2027                 [Page 8]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

   ExternalEffect(A)
       ONLY IF
   AuthorizedDescriptor(A) matches ActualPresentedOperation(A)
   AND CurrentAuthority == VALID
   AND Freshness == VALID
   AND ProtectedState == VALID
   AND EffectuationBoundary == AUTHORIZED

   Otherwise: NO EFFECT

3.  Why Verification Is Required at the Last Effectuation Boundary

   This section explains why verification at the last effectuation
   boundary is structurally important for agentic systems, rather than
   merely being one more place to put an authorization check.

3.1.  Dynamic Agentic Execution Changes the Authorization Problem

   Traditional software is often designed around a comparatively
   predetermined execution path: an application receives an input,
   follows programmed logic, invokes known functions, and produces an
   expected class of output.  Agentic systems can behave differently.
   An AI agent may begin with a high-level objective and then
   dynamically construct or revise a plan, select different tools, react
   to intermediate results, discover new resources, change the sequence
   of operations, choose a different destination, delegate work to
   another agent, retry through another service, or alter its next
   action as new context becomes available.

   The important security consequence is that the action contemplated
   when authority was initially granted may not be identical to the
   action eventually presented for execution:

Das                       Expires 27 March 2027                 [Page 9]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

   Initial User / System Objective
               |
               v
           Step 1 -- Initial Plan
               |
               v
           Step 2 -- Tool Result
               |
               v
           Step 3 -- New Context
               |
               v
           Step 4 -- Re-planning
               |
               v
           Step 5 -- Different Action, Target,
                    Parameter, Route, or Tool
               |
               v
      External Consequence

   This does not necessarily mean the agent is malfunctioning -- dynamic
   replanning may be an intended feature of an agentic system.  The
   security problem is therefore not simply whether the agent had
   legitimate authority when the workflow began; it is whether the
   specific action that emerged from the workflow remains within valid
   authority when that action is about to become consequential.

3.2.  Early Authorization Can Become Semantically Stale

   Suppose an agent is initially authorized at Step 1 to perform a task.
   During subsequent execution the agent receives new information and
   revises its plan, and by Step 5 one or more consequential properties
   -- destination, resource, tool, operation, parameter, recipient,
   route, scope, execution context -- may have changed.  The original
   authorization may still be cryptographically valid, the agent may
   still possess valid credentials, the communication channel may still
   be protected, and the agent identity may still be correct;
   nevertheless, the final operation may no longer correspond to the
   operation or scope originally authorized.  The resulting problem can
   be represented as authority granted at Step 1, followed by the agent
   reasoning, observing, and replanning, followed by an action presented
   at Step 5, with the possibility that A(step1) != A(step5).  This is
   why authorization only at workflow entry does not necessarily
   establish authorization of the final consequence.

Das                       Expires 27 March 2027                [Page 10]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

3.3.  Why Not Verify Only at Every Intermediate Step?

   Intermediate verification remains useful and may substantially reduce
   risk.  This document does not discourage authorization at agent
   invocation, authorization before tool selection, authorization at
   service boundaries, workload identity, transaction-context
   propagation, policy checks, attestation, intermediate approval, or
   other defense-in-depth mechanisms.  However, intermediate checks
   cannot by themselves establish the final effectuation property if the
   action may continue to change after those checks: for any check
   occurring at time t_n, a subsequent component or autonomous decision
   may produce A(t_n+1) != A(t_n).

   The last relevant verification therefore has a special property: no
   additional unverified transformation of the security-relevant action
   is permitted between that verification and the consequence it
   authorizes.  For this reason, execution-finality verification is
   positioned at, or cryptographically coupled to, the last non-
   bypassable effectuation boundary.

3.4.  The Last Irreversible or Externally Consequential Boundary

   The phrase "last irreversible boundary" should be understood
   functionally.  Not every consequence is mathematically or physically
   impossible to reverse -- a later operation may attempt to delete
   transmitted information, restore a database, issue a compensating
   command, reverse a configuration, or recall an output.  However, once
   an external consequence has occurred, restoration cannot necessarily
   recreate the state that existed before effectuation: a transmitted
   secret may already have been copied, a rendered output may already
   have been observed, a radio transmission may already have propagated,
   an API invocation may already have caused downstream actions, a
   physical actuator may already have changed the environment, and a
   persistent state transition may already have been consumed by another
   system.

   Accordingly, this document uses the term *effectuation boundary* to
   mean the last practical and non-bypassable technical point at which
   the proposed action can still be withheld before it produces the
   governed external consequence.  This is the reason for placing final
   verification there.

Das                       Expires 27 March 2027                [Page 11]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

3.5.  The Boundary Observes the Action After Agentic Replanning

   The final effectuation component does not need to understand how the
   AI reasoned, reproduce the model's chain of reasoning, or decide
   whether the agent's plan was intelligent.  Instead, it asks a much
   narrower question: what am I actually being asked to do now?  The
   boundary determines or measures the security-relevant properties of
   the pending operation and compares them against the authority
   applicable to that effect.  For example, the originally authorized
   action might be SEND object=dataset-X destination=service-A
   scope=analysis, while after several autonomous steps the actual
   operation presented at egress might be SEND object=dataset-X
   destination=service-B scope=external-transfer.  The agent may have
   arrived at that result through legitimate autonomous reasoning, but
   that fact does not establish authority for the changed consequence.
   The final boundary therefore evaluates the authorized effect against
   the actual pending effect, rather than attempting to validate every
   reasoning step that produced it.

3.6.  Why the Boundary, Not the AI Model, Holds Final Authority

   A central design objective is to avoid requiring the autonomous
   component to police itself.  If the same agent that plans the action
   also has unrestricted authority to make the action real, then a
   failure, compromise, unexpected replanning decision, or misunderstood
   instruction inside that agent may directly produce the consequence.
   The execution-finality model instead separates reasoning and
   computation, which produce a proposed action, from authority to
   effectuate, which produces the external effect.  The agent remains
   free to reason and re-plan within its computational environment; what
   it does not automatically obtain is the ability to make every
   resulting action consequential.  This produces the invariant that an
   agent may change its plan, but authority does not automatically
   change with the plan -- a new action requires valid authority for
   that new action.

3.7.  Why This Matters Specifically for Multi-Agent Systems

   The problem becomes more significant when multiple agents participate
   in one workflow, for example a human or enterprise delegating to
   Agent A, which delegates to Agent B, which invokes Tool C, which
   forwards to Service D, producing an external consequence.  Each
   component may be correctly authenticated, and each delegation may be
   valid in isolation, but the cumulative workflow may produce an effect
   that was not explicitly represented when the original request began.
   For that reason, the final consequential component may need to answer
   independently: what exact effect am I about to produce, under whose
   authority, for which resource or destination, under which current

Das                       Expires 27 March 2027                [Page 12]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

   state, and is that authority still valid for this effect?  This is an
   interoperability question when different agents, authorization
   systems, tools, and effectuation components are operated by different
   vendors or administrative domains.

3.8.  Architectural Consequence

   The execution-finality model therefore does not treat upstream
   authorization and final verification as interchangeable.  Upstream
   authorization answers "may this workflow or actor proceed?"
   Effectuation-boundary verification answers "may this exact
   consequence occur now?"  Both may be necessary.  The latter is
   positioned at the last non-bypassable effectuation boundary because
   it is the final opportunity to detect divergence between initial
   intent, authorized operation, agentic evolution, and actual pending
   operation before external consequence occurs.  Once the boundary is
   crossed, subsequent governance is generally remediation,
   compensation, attribution, or audit rather than prevention.

   The architectural objective can be summarized as: verify as early as
   useful, but verify the actual consequence as late as necessary --
   immediately before effectuation.  This is the reason the Finality
   Sink belongs at the last boundary.

4.  Why This Is Not Merely an Application-Specific Problem

   A reasonable objection is that effectuation control could be treated
   as an application-design concern: each application could decide what
   operations require approval, how authority is represented, and how a
   final action is checked before execution.  For a closed, single-
   vendor system operating entirely within one administrative trust
   domain, that may be sufficient, and this document does not argue that
   every local application requires a new Internet protocol.

   The standardization question arises when the authorization decision,
   autonomous reasoning, tool invocation, workload execution, and final
   effectuation occur across different implementations, vendors,
   administrative domains, or protocol layers.  A representative
   deployment may resemble:

Das                       Expires 27 March 2027                [Page 13]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

   User / Enterprise Authority
              |
              v
      Authorization Service A
              |
              v
           Agent B
              |
              v
           Agent C
              |
              v
           Tool D
              |
              v
         Service E
              |
              v
    Infrastructure / Effectuation
           Boundary F

   In such an environment, no single application necessarily controls
   the complete path.  The component making the initial authorization
   decision may not be the component that ultimately performs the
   consequential operation: the agent generating the operation may be
   supplied by one vendor, the tool may be operated by another, the
   authorization infrastructure may belong to a third organization, and
   the final network, storage, cloud, telecom, or actuator boundary may
   be controlled by yet another operator.

   The question therefore changes from "can one application internally
   decide whether its own action is allowed?" to "can independently
   implemented systems convey and verify a common understanding of what
   action was authorized and whether the action presented at the final
   boundary still corresponds to that authority?"  That is an
   interoperability question.

4.1.  The Missing Interoperability Semantics

   Without common semantics, each implementation may independently
   define concepts such as action identity, destination binding,
   resource binding, effectuation boundary, sink identity, freshness,
   authorization generation, replay state, single-use authority, state
   consumption, and verification evidence.  Two systems may therefore
   agree that an authorization object is cryptographically valid while
   disagreeing about what that object authorizes.

Das                       Expires 27 March 2027                [Page 14]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

   For example, an authorization service might record "resource-X may be
   released to service-A," which may be interpreted downstream by an
   agent as "resource-X may be processed by workflow-Y," while the final
   infrastructure component sees only SEND resource-X TO endpoint-Z.
   All three components may be functioning correctly according to their
   local implementations.  The failure occurs because there is no shared
   representation connecting authorized intent, authorized action, and
   actual pending effect.  This is precisely the class of problem for
   which protocol standardization can become relevant.

4.2.  Cross-Boundary Authority Cannot Be Assumed from Local Policy

   Application-specific policy is useful when the application controls
   both the decision and the effect.  That assumption weakens in
   distributed and agentic systems.  Consider a chain in which Agent A
   delegates to Agent B, which calls Tool C, which requests Cloud
   Service D, which dispatches to Network, Storage, or Device E.  Each
   transition may cross a trust or administrative boundary: an upstream
   component may know why the action was authorized, while the
   downstream component knows what physical or logical effect is
   actually about to occur.  Neither component necessarily has enough
   information alone.

   The protocol problem is therefore the preservation and verification
   of the relationship between those two facts.  A standardized
   mechanism could allow the downstream component to determine, in an
   interoperable way, which action or action class was authorized, which
   destination or resource was authorized, which boundary may accept the
   authority, whether the authority is still fresh, whether it has
   already been consumed, whether a relevant authorization generation
   has changed, and whether the actual operation presented locally
   corresponds to the authorized operation.  Without common semantics,
   these properties remain bilateral implementation agreements.

4.3.  The Multi-Vendor Case

   The need becomes clearer in heterogeneous deployments, for example an
   enterprise policy system from Vendor A, authorization infrastructure
   from Vendor B, an AI agent platform from Vendor C, an external tool
   from Vendor D, and a cloud, telecom, or storage boundary from Vendor
   E.  Vendor E cannot safely be expected to understand Vendor C's
   private internal representation of agent intent.  Vendor C cannot
   necessarily know Vendor E's local effectuation state.  Vendor B
   cannot predict every later tool selection or autonomous re-planning
   decision.

Das                       Expires 27 March 2027                [Page 15]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

   If these systems are expected to interoperate, they need either
   bilateral proprietary integrations between every relevant pair of
   implementations, or a shared representation and verification model.
   The second case is the potential IETF problem.

4.4.  Agentic Replanning Makes Private Workflow Semantics Insufficient

   The issue is especially relevant to autonomous agents because the
   final operation may not be known when the original workflow begins.
   An agent may receive an objective, query a service, interpret the
   result, revise its plan, choose another tool, and only then generate
   the consequential action.  An authorization system at the start of
   that sequence cannot necessarily encode every concrete property of
   the operation that will emerge at the end of it.  Likewise, the
   component that finally acts cannot safely infer unlimited authority
   merely from the fact that the workflow was authorized at its outset.

   This creates a need to preserve authority across a changing execution
   chain while still allowing the final boundary to evaluate the action
   that actually emerged.  That requirement becomes difficult to solve
   purely inside one application when different parts of the chain are
   operated by different parties.

4.5.  Why Existing Authentication Alone Does Not Solve the
      Interoperability Problem

   Cross-domain systems already possess mechanisms for identifying
   callers and protecting messages, answering questions such as who a
   workload is, who issued a token, whether a message was modified,
   whether a credential is valid, and whether a sender is entitled to
   present it.  The additional question considered here is different:
   does this credential or authority object correspond to this exact
   consequential operation, as observed by this final component, now?

   Answering that question across independently implemented systems
   requires agreement about more than identity.  It may require
   interoperable semantics for action, resource, destination, scope,
   boundary, freshness, state, consumption, and evidence.  This is where
   application-specific authorization can become an Internet
   interoperability problem.

Das                       Expires 27 March 2027                [Page 16]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

4.6.  Standardization Does Not Require Standardizing Business Policy

   A standardized effectuation mechanism would not need to specify
   whether a particular action is legally, commercially, ethically, or
   organizationally permitted; those decisions remain application or
   policy specific.  For example, a protocol might standardize how to
   express action=SEND resource=object-123 destination=service.example
   boundary=egress-7 epoch=42 nonce=N123 without defining whether
   sending object-123 to that destination should be allowed.

   The distinction is between policy, which asks whether an action
   should be permitted, and protocol, which asks how different systems
   can represent, bind, carry, and verify the resulting authority.  IETF
   protocols commonly standardize the latter while leaving the former to
   applications and operators.

4.7.  When This Remains Application-Specific

   This document does not claim that execution-finality semantics
   require Internet standardization in every deployment.  A local
   mechanism may be sufficient where one operator controls the complete
   execution path, all components share one implementation,
   authorization and effectuation occur inside one trusted system, no
   cross-vendor interpretation is required, and no authority context
   crosses administrative boundaries.  In such systems, implementation-
   specific mechanisms may provide the required property.

   Standardization becomes more relevant when authority crosses
   administrative domains, autonomous agents invoke third-party tools,
   multiple vendors participate in one action chain, the final
   effectuation component did not make the original authorization
   decision, effectuation metadata must survive multiple protocol hops,
   independently developed systems must agree on action identity and
   binding, or a downstream verifier must understand authority issued
   upstream.  This defines the proposed interoperability threshold.

4.8.  The IETF Question

   The question for the IETF community is therefore not whether the IETF
   should standardize how every application decides whether an AI action
   is safe.  It is: when independently implemented Internet components
   participate in an autonomous action chain, is there a need for common
   semantics that allow the component controlling the final consequence
   to verify that the exact action presented to it remains within
   currently valid upstream authority?

Das                       Expires 27 March 2027                [Page 17]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

   If existing protocols already provide all of the required semantics,
   this document should identify how they are composed.  If only an
   application profile is required, that may be the appropriate outcome.
   If important cross-vendor semantics remain unspecified, then
   additional standardization may be justified.  The objective is
   therefore not to move application policy into the network; it is to
   determine whether authority-to-effect binding across independently
   implemented systems is an interoperability property that requires
   common protocol semantics.

   *In one sentence:* the application decides what should be allowed;
   standardization may be needed when multiple independent systems must
   agree on what was allowed and verify that the same action is the one
   actually about to become effective.

5.  Mapping the Cross-Vendor Problem to Existing IETF Work

   The cross-vendor problem described in the previous section does not
   necessarily require invention of an entirely new protocol stack.
   Several existing and emerging IETF mechanisms already provide
   important parts of the required architecture, including fine-grained
   authorization, sender constraint, transaction-context propagation,
   workload identity, remote attestation, and agent-to-agent or agent-
   to-tool protocol context.

   The open question is whether those mechanisms, individually or in
   combination, provide sufficient interoperable semantics for the last
   effectuation boundary to determine that the action actually presented
   for effectuation corresponds to the action and authority currently
   permitted.

   This section therefore maps the effectuation-finality requirements to
   relevant IETF work without assuming that any one working group should
   own the complete problem.

5.1.  OAuth

   The OAuth Working Group (Web Authorization Protocol) provides a
   natural starting point for representing and carrying authorization
   across administrative boundaries.

   *Rich Authorization Requests (RFC 9396).* RFC 9396 defines the
   authorization_details parameter for carrying structured, fine-grained
   authorization requirements rather than relying only on coarse scope
   strings.  It allows authorization details associated with particular
   resource or access types to be expressed in structured JSON and
   assigned to an access token [RFC9396].

Das                       Expires 27 March 2027                [Page 18]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

   An execution-finality profile could investigate whether such
   authorization details can represent security-relevant properties such
   as action, resource, destination, permitted scope, effect class, and
   effectuation boundary.  The important question is not whether RAR
   already defines execution finality -- it does not.  The question is
   whether RAR provides an appropriate existing extensibility mechanism
   for expressing the authorized side of an exact-action comparison.
   For example:

   Authorization Infrastructure
           |
           | structured authorization_details
           v
   Agent / Workload
           |
           v
   Downstream Resource

   The remaining issue is how a downstream effectuation component
   determines whether its locally observed pending operation matches
   those authorization details.

   *DPoP (RFC 9449).* DPoP provides sender-constrained OAuth tokens and
   protects against some forms of token theft and replay.  A DPoP proof
   binds cryptographic proof to properties including the HTTP method,
   HTTP URI, proof identifier, and, when applicable, the access-token
   hash [RFC9449].

   This is highly relevant, but it addresses a different property.  DPoP
   can help establish that a presenter possesses the key associated with
   an authorization.  Effectuation-boundary verification may
   additionally require that the consequence a presenter is asking a
   resource to produce is the consequence that was authorized.  RFC 9449
   does not generally bind the DPoP proof to an arbitrary canonical
   representation of an application payload or final physical or logical
   effect.

   An IETF discussion could therefore ask whether certain applications
   need an additional profile or binding mechanism connecting sender-
   constrained authorization to an application-defined canonical action
   descriptor.  This document does not propose modifying DPoP itself as
   a prerequisite.

   *Transaction Tokens.* Transaction Tokens are particularly relevant to
   multi-hop autonomous workflows because authorization and identity
   context may need to survive a chain of intermediary workloads.  A
   representative path may be:

Das                       Expires 27 March 2027                [Page 19]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

   User / Enterprise
          |
          v
   Agent A
          |
          v
   Agent B
          |
          v
   Tool C
          |
          v
   Service D
          |
          v
   Effectuation Boundary

   Transaction context can help downstream services understand the
   origin and security context of the transaction.  The execution-
   finality question is what the last consequential component should do
   with that context -- specifically, should the final component compare
   propagated transaction context against the operation it independently
   observes locally before allowing the consequence?

   If so, additional interoperable semantics may be needed for exact-
   action representation, destination or resource binding, effectuation-
   boundary identity, freshness, single-use or bounded authority,
   consumption state, and finality evidence.  The purpose is not to
   duplicate Transaction Tokens, but to determine whether a final-effect
   verification profile can reuse them.

5.2.  WIMSE

   The Workload Identity in Multi System Environments (WIMSE) Working
   Group is directly relevant to the multi-vendor case.  Its charter
   explicitly addresses workloads distributed across trust boundaries,
   situations without a single centralized controller, propagation of
   context across multi-hop and asynchronous transactions, fine-grained
   least privilege, and cryptographic binding of tokens to workload
   identity and optionally to transactions.  This overlaps strongly with
   the infrastructure through which autonomous actions may travel.

   *Workload identity is necessary but not always sufficient.* WIMSE can
   help answer which workload is making a request, which workload
   originated the transaction, what security context accompanied it, and
   to which workload that context is bound.  The execution-finality
   question adds a further one: what exact consequential operation is
   the final workload or boundary about to perform?

Das                       Expires 27 March 2027                [Page 20]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

   The current WIMSE architecture already discusses propagation and
   augmentation of security context from workload to workload and
   binding context to transactions, which makes WIMSE a potentially
   strong foundation.  However, this document asks whether a further
   interoperable profile is needed when the downstream component must
   compare propagated authority context against the locally observed
   pending effect.

   *Context preservation across agentic replanning.* This becomes
   especially relevant when autonomous intermediaries change the
   execution path, for example when Enterprise Authority delegates to
   Agent B, which delegates to Agent C, which selects Tool D, which
   reaches Infrastructure E.  The original security context may be
   propagated correctly; however, Agent C may legitimately choose a
   different tool or operation after receiving new information.

   The key issue is therefore not necessarily preserving an immutable
   historical plan -- indeed, requiring an immutable plan would conflict
   with the adaptive nature of many agentic systems.  The stronger
   requirement is that any changed consequential action must still be
   evaluated against authority applicable to that changed action before
   effectuation.  WIMSE may provide the workload identity and
   transaction-context plumbing; execution-finality semantics would
   address the final comparison between that propagated context and the
   actual pending consequence.

5.3.  RATS

   Remote ATtestation procedureS (RATS) provides another important
   building block.  RFC 9334 defines an architecture in which an
   Attester produces Evidence, a Verifier appraises that Evidence, and a
   Relying Party consumes Attestation Results when making application-
   specific decisions; Evidence may include measurements, configuration
   information, telemetry, or related claims [RFC9334].  This is
   relevant because an effectuation-boundary verifier may itself need to
   be trusted.

   *Attestation of the verification environment.* A remote party may
   need evidence that the component performing effectuation verification
   is running an expected implementation or configuration.  RATS can
   potentially provide evidence concerning the enforcement component,
   protected execution environment, firmware or software measurement,
   configuration, cryptographic identity, and relevant platform state.
   A policy or authorization service could use Attestation Results as
   one input when deciding whether to trust that component.

Das                       Expires 27 March 2027                [Page 21]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

   *Important limitation.* Attestation should not be described as
   proving that a particular finality decision was correctly executed.
   RFC 9334 provides evidence and appraisal about an Attester's state;
   the Relying Party then uses the Attestation Result as part of its own
   decision process.  Acceptable measured state does not automatically
   prove that a specific pending action was correctly authorized and
   effectuated.

   The two can complement one another: RATS can help establish
   confidence in the environment performing verification, while
   execution-finality semantics determine what that environment must
   verify before permitting the consequence.  A possible composition is:

   RATS
     |
     | establishes evidence about verifier state
     v
   Trusted Effectuation Component
     |
     | compares actual action with authority
     v
   Effect / No Effect

5.4.  AgentProto

   The proposed Agent Communication Protocols (AgentProto) Working Group
   is particularly relevant because its charter is explicitly concerned
   with interoperable user-to-agent, agent-to-agent, and agent-to-tool
   dialogs across platforms, intermediaries, and trust boundaries.  The
   current proposed charter includes propagation of protocol-level
   dialog context across multiple intermediaries and explicitly
   identifies coordination with OAuth, WIMSE, RATS, and related security
   work [AGENTPROTO].  As of September 2026, AgentProto remains in
   initial chartering, so this document does not describe it as an
   established WG deliverable.

   *Dynamic agent interaction.* AgentProto addresses an environment
   where Agent A, Agent B, and Tool C may involve multiple hops,
   changing context, different vendors, and multiple trust boundaries.
   That environment is directly relevant to execution finality because
   the concrete consequential operation may not exist when the dialog
   begins, for example when a user provides an objective, an agent
   obtains information, another agent contributes context, a tool is
   selected, and only then does the final consequential operation
   emerge.

Das                       Expires 27 March 2027                [Page 22]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

   The security requirement therefore cannot always be represented as an
   immutable first-step action.  Instead, the authority context must
   survive the workflow while allowing later actions to be checked when
   they become concrete.

   *Separation of concerns.* A useful separation could be layered as:
   AgentProto maintains the agentic dialog and context; OAuth and WIMSE
   provide identity and authority context; RATS provides evidence about
   trusted components; and effectuation verification determines the
   actual pending consequence and compares it with current authority
   before permitting or denying the effect.

   AgentProto need not define AI reasoning, planning semantics, or tool-
   specific business logic, which are explicitly outside its proposed
   scope.  Likewise, an execution-finality mechanism need not define
   agent communication.  The possible interoperability point is the
   security context that must survive until the consequential operation
   becomes concrete.

5.5.  Why No Single Working Group Solves the Whole Problem

   The cross-vendor effectuation problem spans several distinct
   properties: OAuth addresses fine-grained delegated authorization;
   WIMSE addresses workload identity and multi-hop security context;
   RATS addresses evidence about execution and verification
   environments; AgentProto addresses agentic dialog and context
   propagation; the application or resource carries the semantic meaning
   of the requested operation; and the effectuation boundary determines
   the actual consequence about to occur.  None of these layers should
   be unnecessarily duplicated.

   The architectural question is whether they can be composed so that
   the final effectuation component receives sufficient trustworthy
   information to determine: what authority exists; to which action or
   bounded action class it applies; whether that authority is still
   fresh; which boundary may consume it; what operation is actually
   pending locally; whether the authorized and actual operations match;
   whether the authority has already been consumed; and whether the
   decision can be evidenced if required.

   If existing IETF mechanisms can express all of these properties
   through an agreed profile, a new standalone protocol may not be
   required.  If they cannot, the remaining gap can be identified
   precisely rather than creating a parallel authorization system.

Das                       Expires 27 March 2027                [Page 23]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

5.6.  Possible Standardization Path

   This document therefore does not begin from the assumption that a new
   execution-finality protocol must be created.  A more conservative
   standards path is to identify the effectuation-boundary requirements,
   map each requirement to existing IETF mechanisms, identify properties
   already fully satisfied, identify interoperability gaps, and
   determine whether those gaps require implementation guidance, an
   application profile, extensions to existing work, a reusable security
   object, or a new protocol mechanism.

   This approach avoids duplicating OAuth, WIMSE, RATS, AgentProto, or
   other existing work, and provides a concrete basis for cross-WG
   discussion.

5.7.  Questions for the Relevant IETF Communities

   The mapping above leads to specific questions.

   *For OAuth.* Can existing OAuth mechanisms, including RAR, sender-
   constrained authorization, and Transaction Tokens, express enough
   information for a downstream resource to verify that the actual
   consequential operation remains within the authorization originally
   granted?  If not, what minimal additional application-profile
   semantics are required?

   *For WIMSE.* Can transaction-bound workload context carry the
   security-relevant authority information required across a multi-hop
   agentic workflow while allowing the final workload to validate the
   operation it is actually about to perform?

   *For RATS.* What attestation evidence or results would be appropriate
   for establishing trust in an effectuation-boundary verifier without
   incorrectly treating attestation as proof that each individual
   effectuation decision was correct?

   *For AgentProto.* Which authorization and security context should be
   preserved through an agentic dialog so that a downstream tool or
   effectuation component can make its own final authorization decision?
   Should the dialog protocol merely carry references to such context,
   leaving final-effect semantics to another security building block?

   *Cross-WG question.* Most importantly: can existing IETF mechanisms
   be composed into an interoperable model in which one vendor can
   authorize an action, multiple independent agents or workloads can
   process it, and a different vendor's final effectuation component can
   verify that the action actually presented remains within current
   authority before allowing the consequence?  If the answer is yes, the

Das                       Expires 27 March 2027                [Page 24]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

   useful standards output may be a composition or profile.  If the
   answer is no, the remaining gap should be identified narrowly before
   proposing new protocol machinery.

6.  Terminology and Functional Equivalence

   This document uses several terms to describe functional roles.  These
   terms do not require implementations to use the same names.

     +=====================+=========================================+
     | Term                | Functional interpretation               |
     +=====================+=========================================+
     | Candidate Act       | proposed or pending operation           |
     +---------------------+-----------------------------------------+
     | Non-Effective State | staged, uncommitted, or non-effectuated |
     |                     | operation                               |
     +---------------------+-----------------------------------------+
     | Act Descriptor      | canonical machine-verifiable            |
     |                     | description of the proposed operation   |
     +---------------------+-----------------------------------------+
     | Protected           | protected enforcement point, reference  |
     | Enforcement Domain  | monitor, authority-validation component |
     +---------------------+-----------------------------------------+
     | Scoped Non-Bearer   | narrowly bounded effectuation authority |
     | Capability          |                                         |
     +---------------------+-----------------------------------------+
     | Finality Sink       | effectuation gate, commit boundary,     |
     |                     | egress gate, actuator gate              |
     +---------------------+-----------------------------------------+
     | Sink-Side           | local operation derivation or actual-   |
     | Reconstruction      | action measurement                      |
     +---------------------+-----------------------------------------+
     | Finality Receipt /  | protected verification or enforcement   |
     | LAVR                | evidence                                |
     +---------------------+-----------------------------------------+

               Table 1: Terms and Functional Interpretations

   The underlying architecture distinguishes an upstream machine-
   verifiable description of what is authorized from a sink-side
   description of what is actually presented for effectuation.  The
   sink-side function exists to detect mismatch between those two
   states.

Das                       Expires 27 March 2027                [Page 25]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

7.  Threat Model

   The architecture assumes that an operation may be legitimately
   generated and authorized while one or more subsequent components
   cannot be fully trusted.  The threat model includes the following:

   1.   *Action Substitution.* An authorized operation is replaced by
        another operation before effectuation.

   2.   *Destination Substitution.* The operation type remains the same
        but its destination changes.

   3.   *Parameter Modification.* One or more consequential parameters
        are altered after authorization.

   4.   *Replay.* Previously valid authority is reused to create another
        consequence.

   5.   *Duplicate Effectuation.* A valid action is executed more than
        once.

   6.   *Cross-Sink Use.* Authority intended for one effectuation
        boundary is presented to another.

   7.   *Stale Authority.* Authority was valid when issued but is
        revoked or superseded before effectuation.

   8.   *Policy-Generation Change.* The policy epoch or authorization
        generation changes between approval and execution.

   9.   *Compromised Intermediate Component.* An application, runtime,
        operating system, driver, proxy, network component,
        orchestrator, or controller modifies the operation.

   10.  *Compromised Autonomous Agent.* An AI agent may possess valid
        credentials while generating an unintended or unauthorized
        specific action.

   11.  *TOCTOU.* The action checked at one location differs from the
        action actually used at the effectuation point.

   12.  *Crash and Recovery.* A system verifies authority, performs an
        effect, crashes before recording consumption, and later repeats
        the effect.

   13.  *Rollback.* Protected state is restored to an earlier point at
        which consumed authority appears valid again.

Das                       Expires 27 March 2027                [Page 26]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

   The architecture therefore does not assume that the AI model,
   application, operating system, or every intermediate component
   remains trustworthy after upstream authorization.

8.  Architectural Requirements

   The key words "MUST", "MUST NOT", "SHOULD", "SHOULD NOT", and "MAY"
   in this document are to be interpreted as described in BCP 14
   [RFC2119] [RFC8174] when, and only when, they appear in all capitals,
   as shown here.

   *R1 -- Non-Effectuation Before Verification.* A governed action MUST
   NOT become externally effective until the required effectuation-
   boundary verification succeeds.

   *R2 -- Exact-Act Binding.* Authority MUST be bound to the security-
   relevant attributes needed to distinguish the authorized operation
   from a substituted operation.

   *R3 -- Boundary Binding.* Authority MUST identify the intended
   effectuation boundary, sink, or permitted boundary class where such
   distinction is security-relevant.

   *R4 -- Freshness.* The architecture MUST provide a means to detect
   stale or replayed authority.

   *R5 -- Current Protected State.* Effectuation verification MUST
   account for protected state whose change would invalidate the
   operation.

   *R6 -- Actual-Action Verification.* The effectuation component MUST
   determine, measure, reconstruct, derive, or otherwise verify the
   security-relevant operation actually presented for effectuation.

   *R7 -- Fail-Closed Handling.* If information required for
   effectuation cannot be validated, the governed operation MUST NOT
   become externally effective.

   *R8 -- Consumption and Atomicity.* Where authority is single-use or
   bounded, the system MUST prevent verify → effectuate → crash → reuse
   same authority → effectuate again.  Authority therefore needs to be
   consumed, reserved, invalidated, state-advanced, or otherwise made
   non-reusable before or atomically with effectuation.

   *R9 -- Revocation.* The effectuation component MUST have a mechanism
   for determining whether relevant authority has been revoked, expired,
   superseded, or invalidated.

Das                       Expires 27 March 2027                [Page 27]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

   *R10 -- Sink-Side Independence.* The effectuation component MUST NOT
   depend solely on an upstream assertion of what operation is being
   performed when security-relevant attributes can be determined
   locally.

   *R11 -- Unknown State.* Where a required value cannot be established,
   UNKNOWN != AUTHORIZED.  A missing value MUST NOT automatically be
   interpreted as a valid default.

   *R12 -- Evidence.* Deployments requiring accountability SHOULD
   generate protected evidence of successful or failed effectuation
   verification.

   *R13 -- Least Effectuation Authority.* Authorization SHOULD convey
   only the authority required for the specified action or bounded
   action class.

   *R14 -- No Unguarded Equivalent Path.* A system claiming this
   property MUST NOT expose another unguarded path capable of producing
   the same governed consequence.

9.  Execution-Finality Model

   The model separates four logical functions: COMPUTE, VALIDATE
   AUTHORITY, VERIFY AT EFFECTUATION BOUNDARY, and EFFECTUATE.  A
   conceptual flow is:

Das                       Expires 27 March 2027                [Page 28]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

   +-----------------------------+
   | AI / Agent / Application    |
   +--------------+--------------+
                  |
                  | generates
                  v
   +-----------------------------+
   | Proposed Action             |
   | Status: NON-EFFECTIVE       |
   +--------------+--------------+
                  |
                  v
   +-----------------------------+
   | Protected Authority Check   |
   +--------------+--------------+
                  |
                  | valid bounded authority
                  v
   +-----------------------------+
   | Effectuation Boundary       |
   |                             |
   | Determine Actual Operation  |
   | Verify Binding              |
   | Check Freshness             |
   | Check Current State         |
   +---------+-------------------+
             |
         +---+---+
         |       |
        PASS    FAIL
         |       |
         v       v
      EFFECT   NO EFFECT

   The significant property is not the names of these components.  The
   property is that ExternalEffect depends on
   SuccessfulBoundaryVerification, rather than ExternalEffect
   automatically following UpstreamRequest.

10.  Sink-Side Reconstruction and Actual-Action Verification

Das                       Expires 27 March 2027                [Page 29]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

10.1.  Purpose

   The effectuation component should know what it is actually being
   asked to do.  This document uses *sink-side reconstruction* as
   shorthand for determining the security-relevant properties of the
   pending local operation.  It does not mean rerunning an AI model, and
   it does not require reproducing model reasoning.  It means answering:
   what operation am I actually about to effectuate?

10.2.  Examples

   A network egress boundary could derive:

   action          = TRANSMIT
   destination     = endpoint-A
   payload_digest  = H(payload)
   route_class     = external

   A storage boundary could derive:

   action          = WRITE
   resource        = object-X
   namespace       = tenant-A
   content_digest  = H(content)

   A tool-execution boundary could derive:

   action          = INVOKE
   tool            = service-X
   operation       = operation-Y
   target_resource = object-Z

   An actuator boundary could derive:

   action          = ACTUATE
   device          = controller-X
   command         = OPEN
   parameter       = 42

10.3.  Comparison

   Let D_A be the authorized descriptor and D_S the sink-derived
   descriptor.  Effectuation proceeds only if Match(D_A, D_S,
   CurrentState) == TRUE.  The matching function may use exact equality
   for some fields and constrained or range-based matching for others.

Das                       Expires 27 March 2027                [Page 30]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

11.  Relationship to Existing Internet Security Mechanisms

   This section is deliberately framed as a relationship analysis, not
   as a claim that existing mechanisms are defective.

11.1.  OAuth

   OAuth provides delegated authorization.  Its current charter
   explicitly includes new mechanisms and extensions for automated
   agents acting on behalf of users and across multiple administrative
   domains.  RFC 9396 additionally provides Rich Authorization Requests,
   allowing fine-grained authorization details to be conveyed in OAuth
   messages [RFC9396].  These mechanisms may provide important inputs to
   effectuation-boundary verification.  The question raised by this
   document is narrower: does an interoperable OAuth deployment also
   need standardized semantics for determining that the operation
   actually presented at the final consequence boundary is still the
   operation described by the authorization?  This document does not
   assume the answer must be a new OAuth mechanism -- it asks whether
   the property can be expressed using existing OAuth mechanisms,
   application profiles, transaction context, or additional protocol
   elements.

11.2.  DPoP and Sender-Constrained Authorization

   RFC 9449 defines DPoP as an application-layer proof-of-possession
   mechanism for sender-constraining OAuth tokens [RFC9449].  Sender
   constraint addresses an important question -- is the presenter
   authorized to use this token?  Effectuation-boundary verification may
   additionally require: does this token authorize this exact locally
   observed consequence?  The two properties may be complementary.

11.3.  Transaction Tokens

   The OAuth Transaction Tokens work propagates user identity, workload
   identity, and authorization context through a call chain inside a
   trust domain, and the current draft explicitly discusses compromised
   workloads, parameter modification, and narrowly scoped transaction
   context [OAUTH-TXN].  This makes Transaction Tokens highly relevant.
   An important question for this work is whether Transaction Token
   context can represent and preserve enough information to bind
   authorization to a final consequential operation, and whether the
   final resource or effectuation component needs additional semantics
   for reconstruction, exact-effect comparison, consumption, crash
   consistency, and finality evidence.

Das                       Expires 27 March 2027                [Page 31]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

11.4.  RATS

   RFC 9334 defines an architecture for generating, conveying, and
   evaluating evidence about whether another entity is in an intended
   operating state [RFC9334].  RATS evidence may therefore provide
   useful information about platform state, workload state, trusted
   components, and measurement results.  Execution finality asks a
   related but distinct question: even if this component is in an
   acceptable state, is this exact pending action authorized to become
   consequential now?  A deployment could therefore use RATS evidence as
   an input to an effectuation decision rather than replacing either
   architecture.

11.5.  WIMSE

   The WIMSE working group addresses workload identity and fine-grained
   least-privilege access across multiple service platforms.  Its
   charter includes token solutions that may carry caller identity,
   transaction binding, attestation context, and other information.
   This document is relevant to WIMSE where an identified and authorized
   workload eventually initiates a consequential action after traversing
   multiple services.  The open question is whether workload identity
   plus transaction context is sufficient, or whether some applications
   also require explicit authorized-action-to-actual-effect binding at
   the final resource boundary.

11.6.  COSE

   COSE provides CBOR-based structures for signatures, MACs, encryption,
   keys, and related protected attributes; its working group focuses on
   cryptographic object formats and associated parameters.  COSE may
   therefore be useful as an encoding or protection substrate for an
   execution-finality object.  This document does not propose that COSE
   itself determine effectuation policy.

11.7.  Agent Communication Protocols

   The proposed Agent Communication Protocols work is especially
   relevant.  Its current proposed charter addresses user-to-agent,
   agent-to-agent, and agent-to-tool interactions, protocol context
   propagation, identity, authentication, authorization, security, and
   interoperability across agentic systems, and it explicitly
   distinguishes protocol mechanisms from AI reasoning and behavior
   [AGENTPROTO].  Effectuation-boundary verification may intersect with
   this work where agent dialog leads to an agent decision, which leads
   to a tool invocation, which leads to a real external effect.  The
   relevant question is whether agent-protocol context should carry
   information sufficient for a downstream tool or resource to verify

Das                       Expires 27 March 2027                [Page 32]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

   the consequential action, or whether effectuation finality belongs in
   a separate reusable security building block.  As of September 2026,
   AgentProto is still in initial chartering rather than an established
   adopted WG program, so the relationship should be discussed rather
   than assumed.

12.  Why This Is Relevant to the IETF

   This problem is relevant to the IETF only if it contains an
   interoperability dimension.  A purely local implementation choice
   does not necessarily require protocol standardization.

   The potential interoperability problem appears when different
   administrative domains, vendors, workloads, agents, authorization
   systems, and effectuation components need to communicate what action
   was authorized, for which destination, at which boundary, under which
   freshness state, under which authorization generation, with which
   evidence, and under which consumption semantics.  For example:

   Vendor A Agent
         |
         v
   Vendor B Authorization Service
         |
         v
   Vendor C Workload
         |
         v
   Vendor D Tool
         |
         v
   Vendor E Infrastructure Boundary

   If each component uses private semantics for action identity,
   destination binding, replay state, transaction context, sink
   identity, consumption, or finality evidence, then cross-vendor
   verification becomes difficult.

   The possible IETF problem is therefore not "how should an AI decide
   what to do?"  It is: how can independently implemented Internet
   components interoperably convey and verify the authority needed for
   an autonomous action to become consequential?

13.  Potential Protocol Interaction Model

   This document does not require a new wire format.  A generic
   realization may contain four logical objects.

Das                       Expires 27 March 2027                [Page 33]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

13.1.  Action Descriptor

   ActionDescriptor = {
       action_id,
       action_class,
       source,
       destination,
       resource,
       effect_class,
       scope,
       boundary_id,
       sink_id,
       nonce,
       policy_epoch,
       freshness,
       evidence_refs
   }

   Every field in this descriptor is subject to the canonicalization and
   semantic-equivalence requirements in Section 17.1; a descriptor that
   is cryptographically valid but semantically ambiguous is not valid
   effectuation authority.

13.2.  Authority Decision

   Decision = Validate(
       ActionDescriptor,
       AuthorityContext,
       CurrentState,
       RevocationState,
       PolicyGeneration
   )

13.3.  Bounded Effectuation Authority

   EffectuationAuthority = Protect({
       descriptor_digest,
       scope,
       sink_id,
       boundary_id,
       nonce,
       policy_epoch,
       expiry,
       state_ref,
       evidence_ref
   })

Das                       Expires 27 March 2027                [Page 34]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

13.4.  Effectuation Verification

   Actual = DetermineActualPendingOperation()
   if !VerifyAuthority():
       DENY
   if !Match(Authorized, Actual):
       DENY
   if !Fresh():
       DENY
   if !CurrentStateValid():
       DENY
   if !CorrectBoundary():
       DENY
   ConsumeOrReserveAuthority()
   EFFECTUATE

   The exact representation is intentionally left open for IETF
   discussion.

14.  Failure and Recovery Semantics

14.1.  Verification Failure

   If required verification fails, there is no effect.  Implementations
   may suppress, quarantine, discard, defer, request fresh authority, or
   return an error.

14.2.  Timeout

   Failure to obtain required verification information must not
   automatically become authorization.

14.3.  Crash Before Effectuation

   A crash before final effect should leave the operation non-effective
   where technically possible.

14.4.  Crash After Effectuation

   Recovery must distinguish "effect never happened" from "effect
   happened but accounting state did not complete."  Otherwise duplicate
   effectuation may occur.

14.5.  Partial Effects

   Compound actions may require separate finality boundaries for
   individual irreversible sub-actions.

Das                       Expires 27 March 2027                [Page 35]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

15.  Latency Considerations

   Execution-finality verification does not require every policy
   computation to occur in the latency-critical path.  A deployment may
   separate a control path from a hot effectuation path.  Expensive work
   may occur earlier, including policy evaluation, attestation,
   authority issuance, trust establishment, revocation refresh, and
   credential validation.  The final effectuation path may perform only
   bounded operations such as descriptor matching, nonce verification,
   sink binding, signature or MAC verification, generation checking, and
   protected-state transition.

   The requirement is not to recompute all governance policy at the
   final boundary.  The requirement is that the final boundary must
   possess enough trustworthy information to determine whether the
   pending consequence is currently authorized.

16.  Deployment and Legacy Compatibility

   Execution finality is intended to coexist with existing protocols.  A
   deployment may continue to use TLS, HTTP, OAuth, RAR, DPoP,
   Transaction Tokens, workload identity, RATS evidence, JWT, CWT, COSE,
   and application-specific policy systems; these mechanisms may supply
   inputs to the decision.  The architectural change is that selected
   consequential operations become dependent on successful final
   verification.  Legacy systems may remain upstream provided they
   cannot bypass the governed effectuation path.

17.  Security Considerations

   Implementations need to consider at least: action substitution;
   destination substitution; parameter modification; replay; cross-sink
   presentation; stale authorization; generation mismatch; revocation;
   compromised workloads; confused deputy; TOCTOU; rollback; duplicate
   effectuation; alternate bypass paths; compromised verification state;
   and canonicalization disagreement.

   Particular attention is required for canonicalization.  Two
   components that semantically interpret an action differently could
   compute the same protocol flow while believing different consequences
   are authorized.  Therefore, any standardized descriptor format would
   need unambiguous canonicalization and comparison semantics.

Das                       Expires 27 March 2027                [Page 36]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

17.1.  Canonicalization, Semantic Equivalence, and Descriptor
       Interpretation

   Execution-finality security depends not only on cryptographic
   integrity but also on semantic agreement between the component that
   authorizes an action and the component that verifies the action at
   the effectuation boundary.  Two implementations MUST NOT be allowed
   to interpret the same ActionDescriptor as authorizing materially
   different consequences.

   The relevant invariant is therefore not merely Hash(Descriptor_A) ==
   Hash(Descriptor_B), but Meaning(Descriptor_A) ==
   Meaning(Descriptor_B) for all security-relevant fields.  A descriptor
   that is cryptographically authentic but semantically ambiguous MUST
   NOT be treated as valid effectuation authority.

   *1.  Canonical Representation.* Any field included in a signature,
   MAC, digest, capability binding, or protected authorization object
   MUST have a deterministic canonical representation.  Implementations
   MUST agree on field names, field types, field ordering where ordering
   is significant, byte encoding, string encoding, normalization rules,
   integer representation, timestamp representation, URI representation,
   identifier representation, treatment of optional and duplicate and
   unknown fields, representation of arrays and sets, and canonical
   representation of nested objects.  Implementations MUST NOT
   independently invent normalization rules for security-relevant
   fields.  Where JSON is used, a deterministic canonicalization
   mechanism such as the JSON Canonicalization Scheme (JCS) may be
   appropriate; where CBOR is used, deterministic encoding rules SHOULD
   be used where the encoded form participates in cryptographic binding.
   The specific encoding is not mandated by this architecture, but one
   interoperable encoding profile MUST define one deterministic
   interpretation.

Das                       Expires 27 March 2027                [Page 37]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

   *2.  Unknown, Missing, Empty, and Null Are Different States.* A
   particularly dangerous source of cross-implementation disagreement is
   treating missing, null, empty, unknown, and default as equivalent.
   They MUST NOT be assumed equivalent where the field affects
   effectuation authority: destination = "" is not necessarily
   equivalent to an absent destination, and neither is necessarily
   equivalent to destination = UNKNOWN.  For required security-relevant
   attributes, UNKNOWN MUST NOT be converted into EMPTY, DEFAULT, or
   AUTHORIZED: UNKNOWN != EMPTY, UNKNOWN != DEFAULT, UNKNOWN !=
   AUTHORIZED.  If a required value cannot be determined, the
   effectuation component MUST fail closed unless the applicable profile
   explicitly defines another safe interpretation.  This rule is
   especially important for destination, resource, scope, action class,
   boundary identity, sink identity, policy generation, authorization
   state, provenance, freshness, and protected-state references.

   *3.  Type Confusion MUST Be Prevented.* Implementations MUST NOT
   treat values as equivalent merely because they render similarly: "42"
   must not automatically equal 42, true must not equal "true", and null
   must not be interpreted as false, unless the relevant profile
   explicitly defines such equivalence.  Security-sensitive descriptors
   SHOULD use a schema that assigns one unambiguous type to every field.

   *4.  Duplicate Fields.* Duplicate object members can create severe
   parsing inconsistencies -- different parsers may accept the first,
   accept the last, reject the structure, or preserve both.  An
   execution-finality profile MUST define one behavior; the safest rule
   is that a descriptor containing duplicate security-relevant member
   names MUST be rejected.

   *5.  Unicode and String Normalization.* Textual identifiers can have
   multiple byte representations while appearing visually identical,
   which matters especially for internationalized domain names, user
   identifiers, resource names, filesystem paths, tool names, tenant
   identifiers, and jurisdiction identifiers.  An implementation MUST
   NOT perform undocumented Unicode normalization after authorization
   but before comparison; where a normalized representation is used, the
   normalization method MUST be specified by the profile and applied
   consistently before cryptographic binding.

Das                       Expires 27 March 2027                [Page 38]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

   *6.  URI and Endpoint Canonicalization.* URLs and network endpoints
   are particularly dangerous because multiple textual forms may refer
   to the same resource, while superficially similar forms may refer to
   different resources -- https://example.com/a and https://example.com/
   a/ may have different application semantics, and example.com and
   example.com:443 may or may not be treated equivalently.  Profiles
   using URIs or endpoints MUST define exactly which URI components
   participate in the authorization comparison and how they are
   normalized; implementations MUST NOT rely on visual similarity.

   *7.  Numeric Values and Units.* Numeric ambiguity is especially
   dangerous for autonomous infrastructure: the value 42 has no
   interoperable meaning unless both parties agree whether it represents
   bytes, milliseconds, degrees, percent, volts, or some other unit.
   Where a numeric value affects the consequence, the descriptor SHOULD
   bind the value together with its unit or defined semantic type, for
   example {"parameter":42,"unit":"percent"} rather than
   {"parameter":42}. Floating-point values SHOULD be avoided in
   security-critical equality comparisons unless their representation
   and comparison semantics are explicitly defined.

   *8.  Set Versus Sequence Semantics.* A descriptor must specify
   whether a collection is ordered or unordered -- ["read","write"] must
   not be assumed equivalent to ["write","read"] unless the profile
   defines the field as a set.  Canonicalization must therefore preserve
   semantic collection type, not merely serialized form.

   *9.  Default Values.* Implicit defaults are dangerous across
   implementation boundaries: if Vendor A interprets an absent scope as
   read-only while Vendor B interprets the same absence as unrestricted,
   a valid signature could authorize different consequences in the two
   systems.  For security-relevant fields, defaults MUST either be
   explicitly encoded or be defined identically by the applicable
   interoperability profile; security-sensitive defaults SHOULD
   generally be explicit.

   *10.  Versioning.* ActionDescriptor semantics will evolve, so a
   descriptor SHOULD identify the semantic profile or version under
   which it is interpreted, for example descriptor_profile = "ef-action-
   v1".  A verifier MUST NOT silently interpret a descriptor under a
   newer or different semantic model; if the verifier does not
   understand the declared profile or version, the result SHOULD be
   UNSUPPORTED, and, for a required finality check, NO EFFECT.

   *11.  Unknown Extension Fields.* Extensibility must not create
   downgrade or interpretation attacks.  A verifier encountering an
   unknown field must know whether that field is optional or critical; a
   useful model permits extension fields while allowing security-

Das                       Expires 27 March 2027                [Page 39]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

   critical extensions to be marked as requiring recognition, and if an
   unknown field is marked critical, verification MUST fail.  This
   prevents an authorizing system's essential field from being silently
   ignored by the verifier and producing an unintended effect.

   *12.  Semantic Comparison Must Occur After Canonical Interpretation.*
   The Finality Sink SHOULD NOT simply compare serialized byte strings
   when the locally observed action originates from a different
   representation.  Instead, both the authorized object and the sink-
   observed operation should be mapped into the same defined canonical
   semantic representation before equality or constrained matching is
   performed: D_A = Canonicalize(AuthorizedAction), D_S =
   Canonicalize(SinkObservedAction), Match(D_A, D_S).  The
   canonicalization function MUST itself be part of the security
   specification.

   *13.  Matching Does Not Always Mean Byte Equality.* Some operations
   require bounded authorization rather than exact byte-for-byte
   equality -- an authorized temperature range of 20 to 25 could permit
   a requested temperature of 23, and an authorized destination class of
   *.internal.example may permit a specific internal service.  The
   profile MUST therefore distinguish exact equality from range
   constraint, set membership, prefix constraint, resource class, and
   bounded quantity, with explicit and deterministic comparison
   semantics; a verifier MUST NOT invent broader interpretation than the
   authorizing component expressed.

   *14.  Canonicalization Is Part of the Trusted Computing Base.*
   Canonicalization MUST NOT be treated as cosmetic serialization logic;
   in this architecture it is part of the security boundary.  An
   attacker who can cause the authorizer's meaning of a value to differ
   from the verifier's meaning of the same value may bypass exact-act
   binding even without stealing a key, forging a signature,
   compromising TLS, breaking an HSM, or compromising the AI model.
   This class of vulnerability is therefore equivalent to an
   authorization failure, and the canonicalization and semantic-
   interpretation code SHOULD be kept small, deterministic, testable,
   and suitable for independent conformance testing.

   *15.  Cross-Language Conformance.* Interoperability testing SHOULD
   include implementations written in different programming languages;
   for every normative descriptor test vector, interpretations across
   languages such as Python, Go, Rust, Java, and JavaScript should agree
   for every security-relevant semantic field.  Testing only one
   serializer and parser implementation is insufficient because many
   canonicalization vulnerabilities emerge from differences between
   runtime libraries.

Das                       Expires 27 March 2027                [Page 40]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

   *16.  Negative Test Vectors.* A standard should include adversarial
   vectors in addition to valid examples, covering cases such as
   duplicate keys, unknown fields, unknown critical fields, null versus
   absent, empty versus absent, Unicode equivalents and confusables, URI
   normalization differences, numeric type differences, overflow,
   negative and out-of-range values, unexpected arrays, field and
   nested-object reordering, alternate encodings, stale versions,
   unsupported profiles, malformed timestamps, timezone differences,
   duplicate identifiers, case differences, trailing separators, and
   percent encoding.  Each vector should specify one expected
   interoperable result -- ACCEPT, DENY, INVALID, or UNSUPPORTED -- and
   conforming implementations should not be permitted to disagree about
   a security-relevant interpretation.

   *17.  Canonicalization Failure Is a Finality Failure.* The
   effectuation boundary MUST NOT attempt best-effort interpretation
   when canonicalization fails.  If the descriptor cannot be parsed,
   cannot be canonicalized, cannot be interpreted, is semantically
   ambiguous, contains an unsupported critical field, or produces a
   canonicalization disagreement, the result is NO EFFECT.  This
   preserves the execution-finality invariant.

   *18.  Security Invariant.* The descriptor security property can
   therefore be expressed as VerifyCrypto(D) == TRUE is necessary but
   not sufficient.  Effectuation additionally requires:

   Parse(D) == VALID
   AND Canonicalize(D) == VALID
   AND Semantics(D) == UNDERSTOOD
   AND AuthorizedMeaning(D_A) matches ObservedMeaning(D_S)
   AND CurrentState == VALID

   Only then: EFFECT
   Otherwise: NO EFFECT

   *Security rationale.* A signature establishes integrity over a
   representation.  It does not by itself establish that all
   participating implementations assign the same meaning to that
   representation.  For execution-finality systems, semantic
   interoperability is therefore part of authorization security itself.
   The protocol must ensure not merely that two parties agree on the
   bytes that were signed, but that they agree on the consequence those
   bytes authorize.

   *The security objective is not merely canonical bytes.  It is
   canonical consequence.*

Das                       Expires 27 March 2027                [Page 41]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

18.  Privacy Considerations

   An effectuation descriptor can itself contain sensitive information.
   Potentially sensitive fields include destination, resource, user
   identity, workload identity, operation class, purpose, jurisdiction,
   timing, and authorization context.  Implementations should minimize
   disclosed fields, using techniques such as hashes, opaque references,
   selective disclosure, local derivation, encrypted claims, and
   protected evidence references.

   Execution-finality mechanisms must not create an unnecessary cross-
   domain surveillance record merely in the name of accountability.

19.  IETF Relevance and Possible Venues

   The architecture crosses several existing areas of work.  No single
   group should be assumed to own the complete problem.

   OAuth  Relevant to delegated authority, automated-agent
      authorization, RAR, DPoP, and transaction context.  OAuth's
      current charter explicitly includes complex delegation for
      automated agents.

   WIMSE  Relevant to workload identity, multi-hop execution,
      transaction binding, least privilege, and workload context.

   RATS  Relevant to protected evidence, attestation, measured state,
      and verifier results.

   COSE  Potentially relevant to compact signed effectuation objects and
      cryptographically protected attributes.

   AgentProto  Potentially relevant to agent-to-agent and agent-to-tool
      context, autonomous action workflows, and cross-vendor agent
      interoperability.

   DISPATCH  Potentially relevant as the venue for deciding whether the
      problem belongs in an existing WG, needs a cross-WG design effort,
      requires a new focused venue, or should remain application-
      specific.  The current DISPATCH charter explicitly exists to
      evaluate new work across ART, SEC and non-transport WIT, identify
      overlap with existing protocols, and find an appropriate home
      rather than doing the protocol work itself.

Das                       Expires 27 March 2027                [Page 42]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

20.  Questions for the IETF Community

   This section is deliberately positioned as one of the most visible
   parts of this document.  The posture of this draft is: here is the
   governance-driven technical problem, here are the properties believed
   to matter -- which existing IETF mechanisms already satisfy them,
   which do not, and where, if anywhere, should the remaining work live?
   This gives reviewers something concrete to answer instead of forcing
   them to accept or reject the entire execution-finality architecture.

   *Q1.  Is the problem real?* Does the IETF community agree that there
   is a useful distinction between "this caller is authorized" and "this
   exact consequential operation, as locally presented at the final
   boundary, is currently authorized"?  If not, which existing
   specification already provides the equivalent property?

   *Q2.  Can existing OAuth mechanisms express the required semantics?*
   Can RFC 9396 authorization details, DPoP, Transaction Tokens, token
   exchange, or related OAuth mechanisms fully express exact-act
   binding, sink or boundary binding, local actual-action comparison,
   current protected state, single-use or bounded consumption, and
   crash-consistent effectuation?  If they can, should an application
   profile specify how?

   *Q3.  Is transaction context enough?* Transaction Tokens carry
   authorization context through a call chain.  Should a final resource
   be required to compare that context against the operation it is
   actually about to perform?  If so, is this already implied by normal
   application authorization semantics, or would interoperable behavior
   benefit from specification?

   *Q4.  What should "exact action" mean across implementations?* How
   should independently developed systems agree on the identity of an
   action?  Should standardization define action class, destination,
   resource, effect class, parameter set, and canonical representation,
   or should those semantics remain entirely application-specific?

   *Q5.  Where should reconstruction occur?* Should an effectuation
   component derive the consequential operation from the incoming
   request, local pending state, a protected descriptor, a device or
   controller representation, an application transaction object, or some
   combination?

   *Q6.  Should actual-action measurement be standardized?* Is "what was
   authorized" versus "what is actually about to happen" a general
   protocol abstraction worthy of standardization, or is it inherently
   application-specific?

Das                       Expires 27 March 2027                [Page 43]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

   *Q7.  How should authorization be consumed?* For operations where
   replay creates another consequence, should specifications define
   single-use authority, reservations, epochs, compare-and-commit,
   monotonic state, or idempotency semantics?

   *Q8.  How should crash consistency be represented?* How should an
   interoperable protocol distinguish "authorized but not effectuated"
   from "effectuated but acknowledgement lost" from "authority already
   consumed"?

   *Q9.  How does this relate to RATS?* Should an effectuation component
   use RATS Evidence or Attestation Results to establish its own trusted
   state, upstream protected state, the integrity of an action
   descriptor, or the integrity of the measurement path?  What should
   remain outside RATS?

   *Q10.  How does this relate to AgentProto?* Should an agentic dialog
   carry sufficient context for downstream tools to determine who
   requested the action, on whose behalf, under which authority, for
   which intended effect, and whether that authority remains valid?  Or
   should AgentProto only propagate dialog context while another
   security mechanism controls final effectuation?

   *Q11.  Is a new protocol needed at all?* Could the required property
   be achieved using a profile composed entirely from OAuth, RAR, DPoP,
   Transaction Tokens, WIMSE, RATS, COSE, and existing application
   protocols?  If so, documenting that composition may be preferable to
   inventing another protocol.

   *Q12.  Where is the correct standards boundary?* Possible
   alternatives include: (A) application-specific implementation
   guidance; (B) an OAuth profile; (C) a WIMSE profile; (D) an
   AgentProto security building block; (E) a RATS integration profile;
   (F) a general effectuation-boundary architecture; or (G) a new
   protocol object or exchange.  Which level produces genuine
   interoperability value?

   *Q13.  Is fail-closed behavior always appropriate?* For which classes
   of consequence should unresolved authority result in no effect, and
   where might availability requirements justify different semantics?

   *Q14.  How should multi-agent delegation be treated?* If a human
   delegates to Agent A, which delegates to Agent B, which invokes Tool
   C, which reaches Infrastructure D, which component carries the
   authority chain, and which component ultimately verifies the exact
   effect?

Das                       Expires 27 March 2027                [Page 44]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

   *Q15.  What evidence is actually useful?* Should successful and
   denied effectuation produce standardized receipts?  If so, who
   consumes them, what information is necessary, how is privacy
   preserved, and how are receipts protected from becoming surveillance
   metadata?

21.  Requested Community Feedback

   This document specifically requests feedback on four points:

   1.  Whether the effectuation-boundary problem is sufficiently
       distinct from ordinary authorization to justify explicit
       treatment.

   2.  Whether existing IETF mechanisms already provide all required
       properties when correctly composed.

   3.  If gaps remain, whether they are best addressed through profiles
       of existing work or through additional interoperable objects or
       protocol semantics.

   4.  Which IETF venue, if any, is appropriate for further work.

   The author does not assume that execution finality must become a new
   standalone protocol.  A useful outcome of community review would also
   be a determination that existing IETF mechanisms already provide the
   required properties and that the missing work is merely deployment
   guidance or application profiling.

22.  IANA Considerations

   This document has no IANA actions.

23.  Conclusion

   Autonomous and AI-enabled systems increasingly compress the distance
   between computation and consequence.  The broader governance concern
   has been stated clearly by the UN Secretary-General as the danger of
   technological capability operating without adequate accountability,
   oversight, and transparency.

   The protocol question is narrower.  An authenticated and authorized
   system may still present an action at a consequential boundary that
   differs from the action originally authorized, or that is no longer
   valid under current state.  Execution finality treats this as an
   explicit architectural property: the action actually presented for
   effectuation is verified against currently valid authority before the
   consequence is permitted.

Das                       Expires 27 March 2027                [Page 45]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

   The central invariant is that computation is not authority.

   This document does not presume that a new protocol is required.  It
   asks the IETF community to determine whether the property is already
   covered by existing or emerging work, whether profiles or
   compositions are sufficient, and, if not, what additional
   interoperable mechanisms would be appropriate.

24.  Normative References

   [RFC2119]  Bradner, S., "Key words for use in RFCs to Indicate
              Requirement Levels", BCP 14, RFC 2119, March 1997,
              <https://www.rfc-editor.org/rfc/rfc2119>.

   [RFC8174]  Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC
              2119 Key Words", BCP 14, RFC 8174, May 2017,
              <https://www.rfc-editor.org/rfc/rfc8174>.

25.  Informative References

   [RFC9334]  Birkholz, H., Thaler, D., Richardson, M., Smith, N., and
              W. Pan, "Remote ATtestation procedureS (RATS)
              Architecture", RFC 9334, January 2023,
              <https://www.rfc-editor.org/rfc/rfc9334>.

   [RFC9396]  Lodderstedt, T., Richer, J., and B. Campbell, "OAuth 2.0
              Rich Authorization Requests", RFC 9396, May 2023,
              <https://www.rfc-editor.org/rfc/rfc9396>.

   [RFC9449]  Fett, D., Campbell, B., Bradley, J., Lodderstedt, T.,
              Jones, M., and D. Waite, "OAuth 2.0 Demonstrating Proof of
              Possession (DPoP)", RFC 9449, September 2023,
              <https://www.rfc-editor.org/rfc/rfc9449>.

   [OAUTH-TXN]
              Tulshibagwale, A., "Transaction Tokens", Work in Progress,
              draft-ietf-oauth-transaction-tokens, 2026.

   [AGENTPROTO]
              IETF, "Agent Communication Protocols", Proposed IETF
              Working Group charter, Work in Progress, 2026.

   [UNGA81]   Guterres, A., "Address to the Opening of the General
              Debate of the 81st Session of the General Assembly",
              United Nations, 22 September 2026.

Das                       Expires 27 March 2027                [Page 46]
Internet-Draft   AI Safety at the Effectuation Boundary   September 2026

   [EXEC-FINALITY]
              Das, S., "The Missing Protocol Layer for the Agentic
              Internet: Computation Is Not Authority", Work in Progress,
              draft-das-execution-finality-protocol-layer, 2026.

   [EXEC-HANDLE]
              Das, S., "Possession Is Not Authority: Execution Handle,
              Sink Verification, Atomic Consumption, and Finality
              Receipt", Work in Progress, draft-das-execution-handle,
              2026.

   [TECH-DISCLOSURE]
              Das, S., "The Missing Execution-Finality Layer", Zenodo
              record 22082995, 2026.

Author's Address

   Sangam Das
   Independent
   Balasore
   Odisha
   India
   Email: info@sangamdas.com

Das                       Expires 27 March 2027                [Page 47]