RATS Working Group Kayode Oluwole Elekula Internet-Draft Intended status: Informational 16 November 2026 Expires: 20 May 2027 Attestation of AI Inference Execution Environments Using RATS: Trusted Model Loading, Runtime Binding, and Per-Inference Evidence draft-elekula-rats-model-loading-binding-00 Abstract AI inference services increasingly depend on combinations of model artifacts, inference runtimes, execution environments, hardware accelerators, and security-relevant configuration. A relying party authorizing an AI service to access protected resources may require evidence that an approved model artifact was actually loaded and used by an acceptable inference runtime. Measuring a model artifact and measuring an inference runtime independently does not establish that the measured artifact was the artifact loaded by the measured runtime. Similarly, attesting an execution environment does not, by itself, establish which model processed a particular inference request. This document examines these problems using the Remote ATtestation procedureS (RATS) architecture specified in RFC 9334. It distinguishes execution-instance attestation from per-inference operation evidence and describes a candidate measurement and binding approach using a protected model-loading mechanism, trusted measurement collection, and cryptographically protected evidence. It examines how the Entity Attestation Token (EAT), specified in RFC 9711, and the measured-component information model specified in RFC 10013 may represent relevant measurements and component information. It also examines the relationship between this proposal and existing work on attested AI agent identity and application-layer action evidence. The document does not define a new attestation protocol, EAT profile, registered Claim, Evidence conveyance protocol, or appraisal-policy language. Its objective is to determine whether existing RATS mechanisms can support interoperable appraisal of AI model-to-runtime relationships and to identify any remaining requirements for standardization. Elekula Expires 20 May 2027 [Page 1] Internet-Draft AI Inference Attestation November 2026 The document does not claim that attestation establishes model correctness, safe behavior, or the correctness of generated outputs. 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), its areas, and its working groups. Note that other groups may also distribute working documents as Internet-Drafts. 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." The list of current Internet-Drafts can be accessed at https://www.ietf.org/1id-abstracts.html The list of Internet-Draft Shadow Directories can be accessed at https://www.ietf.org/shadow.html This Internet-Draft is an individual contribution. It does not represent IETF consensus or endorsement by the RATS Working Group. This Internet-Draft will expire on 20 May 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. Code Components extracted from this document must include Revised BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Revised BSD License. Table of Contents 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . 4 1.1. Problem Statement . . . . . . . . . . . . . . . . . . . 5 1.2. Objectives . . . . . . . . . . . . . . . . . . . . . . . 6 1.3. Scope . . . . . . . . . . . . . . . . . . . . . . . . . 7 1.4. RATS Charter Alignment . . . . . . . . . . . . . . . . . 7 Elekula Expires 20 May 2027 [Page 2] Internet-Draft AI Inference Attestation November 2026 1.5. Normative Approach . . . . . . . . . . . . . . . . . . . 7 2. Terminology . . . . . . . . . . . . . . . . . . . . . . . . 8 2.1. AI Inference Execution Instance . . . . . . . . . . . . 8 2.2. Model Artifact . . . . . . . . . . . . . . . . . . . . . 8 2.3. Inference Runtime . . . . . . . . . . . . . . . . . . . 8 2.4. Model-Loading Mechanism . . . . . . . . . . . . . . . . 8 2.5. Model-Loading Relationship . . . . . . . . . . . . . . . 8 2.6. Execution-Instance Attestation . . . . . . . . . . . . . 8 2.7. Per-Inference Evidence . . . . . . . . . . . . . . . . . 8 2.8. Trusted Measurement Agent . . . . . . . . . . . . . . . 8 2.9. Execution Binding . . . . . . . . . . . . . . . . . . . 9 2.10. Inference Operation Identifier . . . . . . . . . . . . 9 3. Concrete Use Case . . . . . . . . . . . . . . . . . . . . . 9 3.1. Deployment Scenario . . . . . . . . . . . . . . . . . . 9 3.2. Security Requirements . . . . . . . . . . . . . . . . . 9 3.3. Assurance Boundaries . . . . . . . . . . . . . . . . . . 10 3.4. Expected Outcome . . . . . . . . . . . . . . . . . . . . 10 4. Attestation Architecture . . . . . . . . . . . . . . . . . . 10 4.1. Attester . . . . . . . . . . . . . . . . . . . . . . . . 10 4.2. Verifier . . . . . . . . . . . . . . . . . . . . . . . . 11 4.3. Relying Party . . . . . . . . . . . . . . . . . . . . . 11 4.4. Endorser . . . . . . . . . . . . . . . . . . . . . . . . 11 4.5. Reference Value Provider . . . . . . . . . . . . . . . . 11 4.6. Trusted Measurement Agent . . . . . . . . . . . . . . . 11 5. Concrete Model-Loading Binding Mechanism . . . . . . . . . . 12 5.1. Fundamental Requirement . . . . . . . . . . . . . . . . 12 5.2. Protected Model-Loading Path . . . . . . . . . . . . . . 12 5.3. Measurement at the Loading Boundary . . . . . . . . . . 13 5.4. Protected Measurement Report . . . . . . . . . . . . . . 13 5.5. Runtime Memory and Accelerator Considerations . . . . . 14 5.6. Limitations . . . . . . . . . . . . . . . . . . . . . . 14 6. Execution-Instance Attestation . . . . . . . . . . . . . . . 15 6.1. Definition . . . . . . . . . . . . . . . . . . . . . . . 15 6.2. Required Measurements . . . . . . . . . . . . . . . . . 15 6.3. Attestation Procedure . . . . . . . . . . . . . . . . . 15 6.4. Freshness . . . . . . . . . . . . . . . . . . . . . . . 15 6.5. State Changes . . . . . . . . . . . . . . . . . . . . . 16 7. Per-Inference Evidence . . . . . . . . . . . . . . . . . . . 16 7.1. Distinction from Execution-Instance Attestation . . . . 16 7.2. Request Association . . . . . . . . . . . . . . . . . . 16 7.3. Trusted Request Dispatcher . . . . . . . . . . . . . . . 17 7.4. Execution Receipt . . . . . . . . . . . . . . . . . . . 17 7.5. Cryptographic Protection . . . . . . . . . . . . . . . . 18 7.6. Per-Inference Freshness . . . . . . . . . . . . . . . . 18 7.7. Limitations . . . . . . . . . . . . . . . . . . . . . . 18 8. Evidence Representation Using Existing RATS Specifications . 19 8.1. RFC 9711 -- Entity Attestation Token . . . . . . . . . . 19 8.2. RFC 10013 -- Measured Component . . . . . . . . . . . . 19 8.3. Composite Representation . . . . . . . . . . . . . . . . 20 8.4. Candidate Information Model . . . . . . . . . . . . . . 20 Elekula Expires 20 May 2027 [Page 3] Internet-Draft AI Inference Attestation November 2026 8.5. Reuse Before Extension . . . . . . . . . . . . . . . . . 20 9. Reference Values and Endorsements . . . . . . . . . . . . . 21 9.1. Reference Values . . . . . . . . . . . . . . . . . . . . 21 9.2. Model Artifact Reference Values . . . . . . . . . . . . 21 9.3. Endorsements . . . . . . . . . . . . . . . . . . . . . . 21 9.4. Applicability . . . . . . . . . . . . . . . . . . . . . 21 10. Verifier Appraisal . . . . . . . . . . . . . . . . . . . . 22 10.1. Appraisal Inputs . . . . . . . . . . . . . . . . . . . 22 10.2. Execution-Instance Appraisal . . . . . . . . . . . . . 22 10.3. Per-Inference Appraisal . . . . . . . . . . . . . . . . 22 10.4. Incomplete Evidence . . . . . . . . . . . . . . . . . . 23 10.5. Attestation Results . . . . . . . . . . . . . . . . . . 23 11. Comparison with Existing RATS Work . . . . . . . . . . . . 23 11.1. RFC 9334 -- RATS Architecture . . . . . . . . . . . . . 23 11.2. RFC 9711 -- Entity Attestation Token . . . . . . . . . 24 11.3. RFC 10013 -- EAT Measured Component . . . . . . . . . . 24 11.4. PTV -- Attested Agent Identity . . . . . . . . . . . . 24 11.5. AEP -- Application-Layer Action Evidence . . . . . . . 25 11.6. Technical Differentiation . . . . . . . . . . . . . . . 25 12. Example End-to-End Deployment . . . . . . . . . . . . . . . 26 12.1. Components . . . . . . . . . . . . . . . . . . . . . . 26 12.2. Execution-Instance Procedure . . . . . . . . . . . . . 27 12.3. Per-Inference Procedure . . . . . . . . . . . . . . . . 27 13. Security Considerations . . . . . . . . . . . . . . . . . . 27 13.1. Evidence Authenticity . . . . . . . . . . . . . . . . . 27 13.2. Measurement Integrity . . . . . . . . . . . . . . . . . 28 13.3. Model Artifact Substitution . . . . . . . . . . . . . . 28 13.4. Model-Loading Manipulation . . . . . . . . . . . . . . 28 13.5. Runtime Modification . . . . . . . . . . . . . . . . . 28 13.6. Accelerator Integrity . . . . . . . . . . . . . . . . . 28 13.7. Execution-Instance Substitution . . . . . . . . . . . . 28 13.8. Per-Inference Evidence Forgery . . . . . . . . . . . . 29 13.9. Replay Attacks . . . . . . . . . . . . . . . . . . . . 29 13.10. Verifier Compromise . . . . . . . . . . . . . . . . . 29 13.11. Model Behavior . . . . . . . . . . . . . . . . . . . . 29 14. Privacy Considerations . . . . . . . . . . . . . . . . . . 29 15. Open Technical Questions . . . . . . . . . . . . . . . . . 30 16. IANA Considerations . . . . . . . . . . . . . . . . . . . . 30 17. Conclusion . . . . . . . . . . . . . . . . . . . . . . . . 31 18. References . . . . . . . . . . . . . . . . . . . . . . . . 31 18.1. Normative References . . . . . . . . . . . . . . . . . 31 18.2. Informative References . . . . . . . . . . . . . . . . 32 Author's Address . . . . . . . . . . . . . . . . . . . . . . . . 32 1. Introduction AI inference services increasingly operate in environments where model execution can affect confidential information, business operations, infrastructure, and security-sensitive decisions. Elekula Expires 20 May 2027 [Page 4] Internet-Draft AI Inference Attestation November 2026 A relying party may therefore require evidence about the execution environment before allowing an AI service to access protected resources. An AI inference deployment may include: * A hardware platform. * A hardware root of trust. * A trusted execution environment, where available. * A boot and operating-system environment. * An inference runtime. * One or more model artifacts. * Model adapters and additional model components. * Runtime and inference configuration. * Hardware accelerators. * Request-dispatch and execution services. Existing model identifiers, provider names, and version strings do not independently establish the integrity of the model artifact actually loaded for execution. Likewise, a cryptographic digest of an artifact establishes the identity of the measured representation but does not, by itself, establish that the artifact was loaded by a particular inference runtime. These distinctions create a specific attestation problem. A relying party may need assurance of the relationship between: 1. An approved model artifact. 2. The artifact measured by a trusted mechanism. 3. The artifact admitted into the inference runtime. 4. The runtime associated with that artifact. 5. The execution instance performing inference. 6. Where required, an individual inference operation processed by that instance. The RATS architecture provides the roles and procedures for evaluating Evidence about system components. This document examines how those procedures may be applied to these AI-specific relationships. 1.1. Problem Statement The principal problem is establishing an independently verifiable relationship between the identity of a model artifact and its actual use by an inference runtime. Elekula Expires 20 May 2027 [Page 5] Internet-Draft AI Inference Attestation November 2026 Consider a deployment in which: * A model artifact has an approved digest. * An inference runtime has an approved digest. * A hardware-backed Attester produces valid Evidence about the execution environment. These facts do not automatically establish that the approved model artifact was loaded by the approved runtime. An implementation that simply combines two independently generated digest values with an ordinary software assertion may still be vulnerable to substitution. For example, compromised software could report the digest of an approved model while loading a different artifact. The required relationship must therefore be supported by a measurement mechanism that observes or controls the model-loading process within an appropriately protected trust boundary. A second problem arises when the relying party needs assurance about an individual inference operation rather than the general state of an execution instance. An attestation result describing an execution instance does not automatically prove that a particular request was processed by that instance. These two problems are related but distinct. 1.2. Objectives The objectives are: 1. Define execution-instance attestation for AI inference environments. 2. Define the distinction between model measurement and model-loading verification. 3. Identify a concrete trusted measurement approach. 4. Examine how the relationship between a model artifact and runtime can be represented in Evidence. 5. Distinguish execution-instance attestation from per-inference evidence. 6. Identify the trust requirements for associating a particular inference request with an attested execution instance. 7. Evaluate existing RATS information models and mechanisms. 8. Compare the proposal with related RATS work. 9. Identify interoperability gaps requiring further investigation. Elekula Expires 20 May 2027 [Page 6] Internet-Draft AI Inference Attestation November 2026 1.3. Scope This document focuses on: * Platform and execution environment measurements. * Inference runtime measurements. * Model artifact measurements. * Trusted model-loading relationships. * Security-relevant runtime configuration. * Execution-instance identification. * Association of inference operations with attested execution instances. It does not define: * A universal AI model identity system. * A new hardware root of trust. * A new attestation protocol. * A new EAT serialization. * A universal inference receipt format. * An appraisal-policy language. * A model-behavior evaluation system. 1.4. RATS Charter Alignment The RATS Working Group charter addresses interoperable attestation of system components and the representation and conveyance of Evidence and Attestation Results. The use case in this document concerns system-component identity, composition, provenance, integrity, configuration, and operational measurements. The proposal therefore investigates an application of existing RATS architectural objectives. It does not assume that AI inference attestation is an adopted RATS Working Group work item. 1.5. Normative Approach The document uses RFC 9334, RFC 9711, and RFC 10013 as the architectural and information-model foundation. Where existing specifications provide suitable representations, they should be reused. Any additional AI-specific semantics should be introduced only where a demonstrable interoperability gap remains. Elekula Expires 20 May 2027 [Page 7] Internet-Draft AI Inference Attestation November 2026 2. Terminology The terminology follows RFC 9334, RFC 9711, and RFC 10013. 2.1. AI Inference Execution Instance An identifiable instance of an AI inference service operating with a particular runtime, model artifact set, and security-relevant configuration. 2.2. Model Artifact A file, collection of files, or other identifiable representation containing model parameters or executable model-related data. 2.3. Inference Runtime The software responsible for loading and executing model artifacts for inference. 2.4. Model-Loading Mechanism The software and associated trusted mechanisms responsible for admitting model artifacts into the inference runtime. 2.5. Model-Loading Relationship The relationship between a measured model artifact and the artifact actually admitted into the inference runtime. 2.6. Execution-Instance Attestation Attestation concerning the identity, integrity, configuration, and relevant component relationships of an AI inference execution instance. 2.7. Per-Inference Evidence Evidence associating an individual inference operation with an identified execution instance and its applicable attested configuration. 2.8. Trusted Measurement Agent A measurement component operating within an appropriately protected trust boundary and responsible for collecting or reporting measurements of relevant components. The security of its Claims depends on its implementation, protection, and relationship to the underlying hardware root of trust. Elekula Expires 20 May 2027 [Page 8] Internet-Draft AI Inference Attestation November 2026 2.9. Execution Binding A verifiable association between an execution instance and its platform, inference runtime, model artifacts, and relevant configuration. 2.10. Inference Operation Identifier An identifier used to correlate an individual inference request and any associated evidence or execution receipt. 3. Concrete Use Case 3.1. Deployment Scenario An organization operates an AI inference service that processes confidential internal information. The service is permitted to access protected resources only when it operates using: * An approved model artifact. * An approved inference runtime. * An acceptable execution environment. * An approved security configuration. The organization requires a Verifier to appraise Evidence before granting access. The organization may also require evidence that individual inference operations were processed by an execution instance satisfying those requirements. 3.2. Security Requirements The deployment requires two distinct assurance levels. Execution-instance requirements 1. The platform is associated with an acceptable attestation identity. 2. The execution environment satisfies the applicable integrity requirements. 3. The inference runtime corresponds to an approved measurement. 4. The model artifact corresponds to an approved measurement. 5. The measurement mechanism establishes the relationship between the measured artifact and the artifact admitted into the runtime. 6. The relevant configuration is acceptable. 7. The Evidence is sufficiently fresh for the intended authorization decision. Elekula Expires 20 May 2027 [Page 9] Internet-Draft AI Inference Attestation November 2026 Per-inference requirements 1. The inference operation has an identifiable request or operation identifier. 2. The request is dispatched to an identified execution instance. 3. The execution instance has an applicable attestation result. 4. The request-dispatch and execution-recording mechanisms are appropriately protected. 5. The resulting evidence is associated with the relevant execution-instance binding. 6. The evidence can be correlated with the intended request without unnecessarily disclosing sensitive inference data. 3.3. Assurance Boundaries The deployment must distinguish between properties established by hardware-backed measurements and properties asserted by ordinary application software. A hardware root of trust can protect keys and measurements, but it does not automatically measure every software component or observe every model-loading operation. The trusted measurement and loading mechanisms must be identified explicitly. 3.4. Expected Outcome The Verifier produces Attestation Results describing the properties successfully appraised. The Relying Party uses those results to authorize access or require additional evidence. Where per-inference assurance is required, the Relying Party additionally evaluates the evidence associating the individual request with the attested execution instance. 4. Attestation Architecture The architecture remains based on RFC 9334. 4.1. Attester The Attester produces Evidence about the AI inference execution environment. The Attester may include a hardware-backed Attesting Environment, a protected measurement agent, and the Target Environment being measured. Elekula Expires 20 May 2027 [Page 10] Internet-Draft AI Inference Attestation November 2026 4.2. Verifier The Verifier appraises Evidence using: * Reference Values. * Endorsements. * Appraisal Policy for Evidence. * Applicable trust anchors. * Freshness information. It produces Attestation Results. 4.3. Relying Party The Relying Party consumes Attestation Results and applies its application-specific authorization policy. 4.4. Endorser The Endorser provides relevant information about the capabilities and identity of the Attester. 4.5. Reference Value Provider The Reference Value Provider supplies acceptable measurements for the relevant components. 4.6. Trusted Measurement Agent The trusted measurement agent is a deployment component, not a new RATS architectural role. It may operate within a TEE or another protected environment. Its responsibilities may include: * Obtaining the inference runtime measurement. * Observing the model-loading process. * Measuring the model representation admitted into the runtime. * Associating measurements with an execution-instance identifier. * Producing integrity-protected measurement information. * Protecting measurement provenance. The mechanism must be protected against modification by the components whose properties it reports. A software component that merely reads a file digest and signs an assertion with an otherwise accessible application key does not necessarily provide independent measurement assurance. Elekula Expires 20 May 2027 [Page 11] Internet-Draft AI Inference Attestation November 2026 5. Concrete Model-Loading Binding Mechanism This section defines the candidate technical approach. It is an architectural mechanism for evaluation, not a newly standardized protocol. 5.1. Fundamental Requirement The measurement mechanism must establish that the model artifact measured is the artifact admitted into the inference runtime. The measurement must be associated with the runtime and execution instance that consume the artifact. The mechanism must not rely exclusively on an ordinary application-level assertion. 5.2. Protected Model-Loading Path A candidate implementation uses the following components: 1. A hardware-backed execution environment. 2. A protected measurement agent. 3. A measured inference runtime. 4. A model-loading mechanism operating within the protected trust boundary. 5. An attestation key protected by an appropriate hardware-backed mechanism. The measurement agent observes or controls the model-loading operation. Where the model is loaded from a file or artifact package, the mechanism verifies and measures the exact artifact representation consumed by the runtime. Where transformations occur before loading, the measurement definition must identify the relevant representation. For example, if a model is quantized, decompressed, assembled from shards, or combined with adapters, the deployment must define whether the relevant measurement represents: * The original packaged artifact. * The transformed artifact. * The final assembled representation. * A manifest covering the constituent artifacts and transformations. The chosen representation must be unambiguous. Elekula Expires 20 May 2027 [Page 12] Internet-Draft AI Inference Attestation November 2026 5.3. Measurement at the Loading Boundary A candidate implementation performs the following operations: 1. The protected environment establishes the identity of the inference runtime. 2. The model-loading mechanism receives the model artifact. 3. The loading mechanism verifies the artifact against an authenticated manifest or applicable integrity requirements. 4. The trusted measurement agent measures the representation actually admitted into the runtime. 5. The agent associates the resulting measurement with the runtime identity. 6. The agent associates both measurements with an execution-instance identifier. 7. The resulting measurement information is protected by an attestation mechanism rooted in the relevant hardware-backed trust mechanism. This approach creates a stronger relationship than independently collecting a runtime digest and model digest. The critical property is that the measurement mechanism observes or controls the loading boundary. 5.4. Protected Measurement Report The conceptual information to be bound includes: +----------------------------+------------------------------------+ | Information | Purpose | +----------------------------+------------------------------------+ | Execution-instance | Identifies the measured instance | | identifier | | +----------------------------+------------------------------------+ | Runtime measurement | Identifies the measured inference | | | runtime | +----------------------------+------------------------------------+ | Model artifact measurement | Identifies the artifact admitted | | | for execution | +----------------------------+------------------------------------+ | Configuration measurement | Identifies relevant security | | | configuration | +----------------------------+------------------------------------+ | Loading-event identifier | Identifies the applicable loading | | | operation | +----------------------------+------------------------------------+ | Measurement epoch | Identifies the measurement period | +----------------------------+------------------------------------+ | Measurement provenance | Identifies the trusted measurement | Elekula Expires 20 May 2027 [Page 13] Internet-Draft AI Inference Attestation November 2026 | | mechanism | +----------------------------+------------------------------------+ | Freshness context | Associates Evidence with an | | | appraisal challenge | +----------------------------+------------------------------------+ These are information requirements, not registered EAT Claims. The exact representation should be selected through evaluation of RFC 9711, RFC 10013, and applicable EAT profile mechanisms. 5.5. Runtime Memory and Accelerator Considerations Measuring an artifact before loading does not necessarily establish the contents of runtime memory after subsequent modification. Where the runtime transforms or copies model data into memory, the deployment must establish how the measurement relates to the representation actually used. Where model data is transferred to an accelerator, a host-side measurement alone does not prove the integrity of accelerator-resident data. Additional mechanisms may be required, such as: * Authenticated transfer. * Device-backed attestation. * Protected memory. * Verified accelerator loading. * Runtime integrity monitoring. If these mechanisms are unavailable, the Evidence must state that accelerator-resident integrity is outside the attested boundary. 5.6. Limitations The proposed mechanism depends on the capabilities of the underlying platform. Not every TEE can independently inspect all model-loading operations or accelerator memory. A trusted measurement agent must therefore be selected and evaluated according to the actual hardware and runtime architecture. No universal hardware capability is assumed. Elekula Expires 20 May 2027 [Page 14] Internet-Draft AI Inference Attestation November 2026 6. Execution-Instance Attestation 6.1. Definition Execution-instance attestation evaluates the identity and integrity of an AI inference execution instance. It concerns the environment in which inference may take place, rather than proof of a particular inference request. 6.2. Required Measurements Potential measurements include: * Platform identity. * Boot measurements. * Execution environment measurements. * Inference runtime measurements. * Model artifact measurements. * Model-loading relationship information. * Security-relevant configuration. * Execution-instance identifier. 6.3. Attestation Procedure A candidate procedure is: 1. The execution environment starts. 2. The platform establishes its applicable attestation state. 3. The inference runtime is measured. 4. The protected model-loading mechanism measures the model representation admitted into the runtime. 5. The trusted measurement agent associates the measurements with the execution instance. 6. Evidence is generated. 7. The Evidence is conveyed to the Verifier. 8. The Verifier appraises the Evidence using applicable Reference Values and Endorsements. 9. The Verifier produces Attestation Results. 10. The Relying Party evaluates the results. 6.4. Freshness Freshness may be established through an appropriate challenge-response mechanism, including an EAT nonce where applicable. A nonce establishes freshness of the attestation exchange when correctly implemented. Elekula Expires 20 May 2027 [Page 15] Internet-Draft AI Inference Attestation November 2026 It does not independently establish which model processed an inference request. 6.5. State Changes The deployment must define which changes invalidate or require renewed attestation. Examples include: * Model artifact replacement. * Runtime replacement. * Configuration modification. * Execution environment restart. * Model adapter replacement. * Changes to the model-loading mechanism. A previously acceptable Attestation Result must not automatically be treated as proof of the current state after a relevant change. 7. Per-Inference Evidence 7.1. Distinction from Execution-Instance Attestation Execution-instance attestation establishes properties of an inference execution environment. Per-inference evidence associates an individual inference operation with that environment. The two are not interchangeable. A fresh execution-instance attestation does not automatically prove that a particular request was processed by the attested instance. 7.2. Request Association A deployment requiring per-inference assurance needs a trusted mechanism at the request-dispatch or execution boundary. The mechanism should establish: 1. The identity of the incoming inference operation. 2. The execution instance selected to process it. 3. The applicable model-to-runtime binding. 4. The association between the request and the execution event. 5. The integrity of the resulting evidence. Elekula Expires 20 May 2027 [Page 16] Internet-Draft AI Inference Attestation November 2026 7.3. Trusted Request Dispatcher A candidate implementation uses a protected request dispatcher or execution gateway. The dispatcher: * Receives an inference request. * Assigns or validates an operation identifier. * Selects an execution instance. * Verifies that the instance has an applicable acceptable attestation state. * Dispatches the request through the controlled execution path. * Records the association between the request and the selected execution instance. The dispatcher must be inside the relevant trust boundary or be independently attested and authenticated. If an ordinary application component can bypass the dispatcher or fabricate its records, the resulting evidence does not independently establish request-to-execution association. 7.4. Execution Receipt A candidate execution receipt may contain the following information: +----------------------------+------------------------------------+ | Information | Purpose | +----------------------------+------------------------------------+ | Operation identifier | Identifies the inference operation | +----------------------------+------------------------------------+ | Execution-instance | Identifies the execution | | identifier | environment | +----------------------------+------------------------------------+ | Model binding reference | Identifies the applicable | | | model-to-runtime binding | +----------------------------+------------------------------------+ | Request context | Associates the request with the | | | operation | +----------------------------+------------------------------------+ | Execution sequence | Supports ordering and correlation | +----------------------------+------------------------------------+ | Measurement epoch | Identifies the applicable | | | execution state | +----------------------------+------------------------------------+ | Output digest, where | Identifies a particular output | | appropriate | representation | +----------------------------+------------------------------------+ | Receipt provenance | Identifies the protected | Elekula Expires 20 May 2027 [Page 17] Internet-Draft AI Inference Attestation November 2026 | | receipt-producing mechanism | +----------------------------+------------------------------------+ Sensitive request inputs and outputs need not be disclosed merely to establish this association. An output digest may be included when the deployment requires output correlation, subject to privacy and confidentiality considerations. 7.5. Cryptographic Protection The receipt must be protected using a mechanism appropriate to the deployment's trust model. Where the receipt is produced inside a protected environment, its integrity may be bound to the relevant attestation identity or protected key. However, a signature alone proves only that a particular key signed the receipt. It does not prove that the reported execution actually occurred unless the receipt-producing mechanism is sufficiently trusted and its relationship to the execution path is established. 7.6. Per-Inference Freshness A deployment may use a unique operation identifier, a nonce, a sequence number, or an appropriate combination. A challenge nonce may be included in a per-inference evidence exchange when the relying party requires challenge-specific freshness. However, freshness and execution association remain separate properties. The operation identifier correlates the request and receipt. The trusted dispatch and execution mechanisms establish the relationship to the execution instance. The model-loading binding establishes the applicable model and runtime identity. No single nonce substitutes for these mechanisms. 7.7. Limitations Per-inference evidence does not automatically prove: Elekula Expires 20 May 2027 [Page 18] Internet-Draft AI Inference Attestation November 2026 * That the model reasoned correctly. * That the generated output is accurate. * That the model followed every application policy. * That the output caused a particular external effect. * That an external tool completed an operation successfully. Those properties require separate evidence and evaluation mechanisms. 8. Evidence Representation Using Existing RATS Specifications 8.1. RFC 9711 -- Entity Attestation Token RFC 9711 provides an attested Claims framework for describing the state and characteristics of an entity. Its measurement and submodule mechanisms are relevant to the representation of multiple components in an AI inference environment. The deployment should evaluate whether an EAT Claims Set can represent: * Platform measurements. * Runtime measurements. * Model artifact measurements. * Configuration information. * Relevant component relationships. * Freshness information. An EAT profile may be necessary to define interoperable semantics for the AI-specific information. This document does not define that profile. 8.2. RFC 10013 -- Measured Component RFC 10013 defines an information model and data models for measured components. It provides a representation for identifying a component and describing its measured state, including digested or raw measurement values. This is directly relevant to model artifact and inference runtime measurements. However, RFC 10013 does not itself define: * A trusted AI model-loading mechanism. * A universal model-loading event. * A standardized relationship between a model artifact and a Elekula Expires 20 May 2027 [Page 19] Internet-Draft AI Inference Attestation November 2026 runtime. * Proof that an independently reported digest corresponds to the artifact actually loaded. * Per-inference request association. Those properties depend on the measurement mechanism and any applicable profile or additional semantics. 8.3. Composite Representation Where an AI deployment contains multiple measured components, existing EAT mechanisms should be evaluated for representing their relationships. A deployment may use appropriate submodules or related Claims Sets to convey component information. The relationship between a model artifact and a runtime must nevertheless be clearly established. Merely placing two component measurements in the same token does not prove that one was loaded by the other. 8.4. Candidate Information Model The following conceptual relationship is relevant: Execution Instance * Platform measurement. * Runtime measurement. * Model artifact measurement. * Configuration measurement. * Model-loading relationship. * Measurement provenance. * Applicable measurement epoch. Where per-inference assurance is required, additional operation evidence may reference the execution-instance identifier and the applicable model-to-runtime binding. These are information requirements for technical evaluation, not a proposed registered Claim vocabulary. 8.5. Reuse Before Extension An implementation should first determine whether existing EAT Claims, measured-component structures, submodules, and profiles can express the required information. Elekula Expires 20 May 2027 [Page 20] Internet-Draft AI Inference Attestation November 2026 An additional AI-specific Claim or profile should only be proposed where existing semantics cannot support interoperable interpretation. 9. Reference Values and Endorsements 9.1. Reference Values Reference Values may identify acceptable measurements for: * Platform firmware. * Boot components. * Operating-system components. * Inference runtimes. * Model artifacts. * Security-relevant configurations. * Model-loading mechanisms. 9.2. Model Artifact Reference Values A model artifact digest may be distributed as a Reference Value. The Verifier may compare the measured artifact with the approved value. However, matching a Reference Value does not independently establish that the artifact was loaded by the runtime. That relationship must be supported by the measurement mechanism. 9.3. Endorsements Endorsements may establish relevant information about: * Hardware attestation capabilities. * Protected keys. * Trusted execution capabilities. * Measurement mechanisms. * Platform identity. An Endorsement about a hardware platform does not automatically endorse every application component operating on that platform. 9.4. Applicability Reference Values and Endorsements must be evaluated within their intended component, platform, version, and deployment context. Elekula Expires 20 May 2027 [Page 21] Internet-Draft AI Inference Attestation November 2026 10. Verifier Appraisal 10.1. Appraisal Inputs The Verifier may receive: * Platform Evidence. * Runtime measurements. * Model artifact measurements. * Configuration measurements. * Model-loading relationship information. * Measurement provenance. * Freshness information. * Endorsements. * Reference Values. Where per-inference evidence is required, the Verifier may additionally receive an execution receipt and the applicable execution-instance Attestation Result. 10.2. Execution-Instance Appraisal The Verifier evaluates: 1. Evidence authenticity. 2. Attester identity. 3. Evidence freshness. 4. Platform measurements. 5. Runtime measurements. 6. Model artifact measurements. 7. Model-loading relationship provenance. 8. Configuration measurements. 9. Component relationships. 10. Applicable Reference Values and Endorsements. 10.3. Per-Inference Appraisal Where per-inference assurance is required, the Verifier additionally evaluates: 1. The authenticity of the execution receipt. 2. The operation identifier. 3. The execution-instance identifier. 4. The applicable model-to-runtime binding. 5. The association between the request and the execution instance. 6. The trustworthiness of the request-dispatch mechanism. 7. The relevant freshness and ordering information. The Verifier must not infer request-to-execution association merely from the existence of an acceptable execution-instance Attestation Elekula Expires 20 May 2027 [Page 22] Internet-Draft AI Inference Attestation November 2026 Result. 10.4. Incomplete Evidence The Verifier should distinguish between: * Successful appraisal. * Failed appraisal. * Incomplete Evidence. * Unauthenticated Evidence. * Missing Reference Values. * Unverified model-loading relationships. * Unverified request-to-execution association. If the model artifact and runtime are independently measured but their loading relationship cannot be established, the Verifier should not report that the model-to-runtime binding has been verified. 10.5. Attestation Results The Verifier produces Attestation Results according to the applicable RATS information model. The result should communicate which properties have been appraised and which remain outside the available assurance boundary. 11. Comparison with Existing RATS Work The proposal is intended to complement existing RATS specifications and related Internet-Drafts. It does not claim that AI attestation, agent identity, or application-layer evidence is a new concept. Its focus is a specific measurement and association problem. 11.1. RFC 9334 -- RATS Architecture RFC 9334 defines the architectural roles and conceptual artifacts used by this proposal. It provides the Attester, Verifier, Relying Party, Evidence, Endorsements, Reference Values, and Attestation Results. The present document applies that architecture to AI inference component relationships. It does not replace or modify the RATS architecture. Elekula Expires 20 May 2027 [Page 23] Internet-Draft AI Inference Attestation November 2026 11.2. RFC 9711 -- Entity Attestation Token RFC 9711 provides the attested Claims framework. Its measurement and submodule facilities provide mechanisms that may represent relevant AI execution components. The present proposal investigates whether an interoperable profile or additional relationship semantics are necessary to express the model-loading relationship. It does not propose a replacement for EAT. 11.3. RFC 10013 -- EAT Measured Component RFC 10013 provides an information model and data models for measured components. It can represent component identity and measurements, including cryptographic digests. Its scope is measurement representation. It does not define how an AI model-loading mechanism establishes that a measured artifact is the one admitted into an inference runtime. The proposed investigation concerns the trusted measurement process and the semantics of the relationship between the measured components. 11.4. PTV -- Attested Agent Identity The individual Internet-Draft, The Prove-Transform-Verify (PTV) Protocol for Attested Agent Identity, addresses hardware-anchored AI agent identity. Its current design binds attestation to a hardware attestation key, uses challenge nonces for freshness, and allows appraisal of configuration hashes against endorsed Reference Values. It also explicitly distinguishes identity-binding integrity from behavioral continuity. Its current threat model identifies model-weight replacement and fine-tuning after enrollment among the limitations it does not address. The present document is concerned with a more specific execution-level question: Elekula Expires 20 May 2027 [Page 24] Internet-Draft AI Inference Attestation November 2026 Can the measurement mechanism establish that the model representation actually admitted into the inference runtime corresponds to the approved measured artifact, and can that relationship be associated with the execution instance? The proposal does not seek to replace PTV's agent identity or challenge-response mechanisms. Where appropriate, those mechanisms may be complementary. PTV provides relevant identity and freshness mechanisms; the present investigation focuses on the model-loading boundary and the measurement provenance needed to establish model-to-runtime association. 11.5. AEP -- Application-Layer Action Evidence The individual Internet-Draft, Composing Application-Layer Action Evidence with Remote Attestation Procedures, examines how signed application-layer records of automated actions may be associated with platform Evidence. Its focus includes action identifiers, authority, action descriptions, outcomes, and the association of application-layer records with an attested platform state. It explicitly recognizes that a signed action record does not independently establish the truth of every reported action or outcome. The present document addresses a different relationship. It investigates whether the model artifact actually loaded by an inference runtime can be independently measured and associated with the execution instance. AEP action evidence may identify what an application reports it did. Model-to-runtime binding evidence concerns which measured model representation was admitted into the execution environment. These mechanisms may be complementary, but they do not establish identical properties. 11.6. Technical Differentiation The following table summarizes the distinction. +---------------+-------------------------+-----------------------+ | Specification | Principal focus | Relationship to this | Elekula Expires 20 May 2027 [Page 25] Internet-Draft AI Inference Attestation November 2026 | or proposal | | document | +---------------+-------------------------+-----------------------+ | RFC 9334 | RATS architecture | Architectural | | | | foundation | +---------------+-------------------------+-----------------------+ | RFC 9711 | Attested Claims | Evidence | | | | representation | +---------------+-------------------------+-----------------------+ | RFC 10013 | Measured components | Component measurement | | | | representation | +---------------+-------------------------+-----------------------+ | PTV | Agent identity and | Complementary | | | configuration binding | identity and | | | | freshness mechanisms | +---------------+-------------------------+-----------------------+ | AEP | Application-layer | Complementary | | | action evidence | operation and action | | | | records | +---------------+-------------------------+-----------------------+ | This proposal | Trusted model-loading | Specific AI inference | | | relationship and | measurement problem | | | execution association | | +---------------+-------------------------+-----------------------+ The proposed contribution is not a claim that existing specifications cannot carry the relevant measurements. It is an investigation into whether they provide sufficient interoperable semantics and trustworthy measurement procedures to establish the required relationships. 12. Example End-to-End Deployment 12.1. Components The example deployment contains: * An AI inference server. * A hardware-backed execution environment. * A trusted measurement agent. * An inference runtime. * An approved model artifact. * A protected model-loading mechanism. * A Verifier. * A Reference Value Provider. * An Endorser. * A Relying Party. Where individual inference operations require separate evidence, a protected request dispatcher is also included. Elekula Expires 20 May 2027 [Page 26] Internet-Draft AI Inference Attestation November 2026 12.2. Execution-Instance Procedure 1. The platform establishes its applicable attestation state. 2. The protected execution environment starts. 3. The inference runtime is measured. 4. The model-loading mechanism receives the approved artifact. 5. The protected measurement agent measures the representation admitted into the runtime. 6. The measurement agent associates the model and runtime measurements with the execution instance. 7. The Attester generates Evidence. 8. The Evidence is conveyed to the Verifier. 9. The Verifier obtains applicable Reference Values and Endorsements. 10. The Verifier appraises the Evidence. 11. The Verifier produces Attestation Results. 12. The Relying Party evaluates the results. 12.3. Per-Inference Procedure Where the deployment requires per-inference evidence: 1. A request reaches the protected request dispatcher. 2. The dispatcher assigns or validates an operation identifier. 3. The dispatcher selects an execution instance. 4. The dispatcher checks the applicable attestation state. 5. The request is forwarded through the controlled execution path. 6. The trusted execution mechanism records the association between the operation and the execution instance. 7. An execution receipt is produced. 8. The receipt is protected and associated with the relevant model-to-runtime binding. 9. The Relying Party or Verifier evaluates the receipt according to the deployment requirements. The procedure does not establish the correctness of the inference result. It establishes only the properties supported by the measurement, execution, and receipt mechanisms. 13. Security Considerations 13.1. Evidence Authenticity Evidence must be protected against forgery, substitution, and unauthorized modification. Elekula Expires 20 May 2027 [Page 27] Internet-Draft AI Inference Attestation November 2026 13.2. Measurement Integrity The measurement mechanism must be protected against modification by the components whose integrity it reports. 13.3. Model Artifact Substitution An attacker may replace an approved model artifact with another artifact. The deployment should measure the representation actually admitted into the inference runtime rather than relying exclusively on a separately reported package digest. 13.4. Model-Loading Manipulation A compromised loading mechanism may report an approved digest while admitting a different artifact. The trusted measurement boundary must include the mechanism responsible for observing or controlling the relevant loading operation. 13.5. Runtime Modification A runtime measurement at startup does not necessarily establish continued integrity. The deployment should consider runtime integrity monitoring and re-attestation following relevant changes. 13.6. Accelerator Integrity Host-side measurements do not automatically establish the integrity of model data resident in a separate accelerator. Where accelerator integrity is required, the deployment needs appropriate device attestation, protected loading, or other verifiable mechanisms. 13.7. Execution-Instance Substitution An attacker may attempt to associate a valid model measurement with a different runtime or execution instance. The relevant component relationships must be integrity-protected and included in the appraisal context. Elekula Expires 20 May 2027 [Page 28] Internet-Draft AI Inference Attestation November 2026 13.8. Per-Inference Evidence Forgery An ordinary application component may fabricate an execution receipt. The receipt-producing mechanism must be appropriately protected and associated with the attested execution path. 13.9. Replay Attacks Previously issued Evidence or execution receipts may be replayed. Freshness, operation identifiers, sequence information, and applicable challenge-response mechanisms should be used according to the deployment threat model. 13.10. Verifier Compromise A compromised Verifier may produce misleading Attestation Results. The Relying Party must consider the trust relationship with the Verifier. 13.11. Model Behavior Attestation does not establish model correctness, fairness, safety, accuracy, or policy compliance. These properties require separate evaluation mechanisms. 14. Privacy Considerations Evidence may disclose: * Hardware identifiers. * Model deployment information. * Runtime versions. * Configuration details. * Execution-instance identifiers. * Inference operation metadata. Per-inference evidence may be particularly sensitive because it can reveal service usage patterns or operational activity. Implementations should consider: * Data minimization. * Access control. * Retention limits. * Confidentiality protection. * Pseudonymous operation identifiers. Elekula Expires 20 May 2027 [Page 29] Internet-Draft AI Inference Attestation November 2026 * Avoidance of unnecessary inference-input disclosure. * Avoidance of unnecessary output disclosure. * Protection against correlation through persistent identifiers. An output digest or request digest should not automatically be treated as privacy-preserving. Low-entropy values may be susceptible to guessing or correlation. 15. Open Technical Questions The following questions require further investigation. * Which existing hardware-backed measurement mechanisms can reliably observe the model-loading boundary? * Can an EAT profile represent the required model-to-runtime relationship without introducing new Claims? * Can RFC 10013 measured-component representations adequately identify the relevant model and runtime measurements? * How should the measurement represent quantized, sharded, or dynamically assembled models? * How should post-loading transformations be measured? * How can accelerator-resident model integrity be established? * What minimum trusted computing base is necessary for model-loading verification? * How should an execution-instance identifier be associated with component measurements? * How should runtime or model changes affect existing Attestation Results? * What mechanisms can reliably associate an individual inference request with an attested execution instance? * Can existing RATS Evidence and Attestation Result representations support this association? * How should per-inference evidence be correlated with execution-instance attestation? * Which existing specifications already provide sufficient semantics for these relationships? * Is a new EAT profile or narrowly scoped information-model extension actually necessary? These questions should be resolved through technical review, implementation experience, and evaluation of existing RATS specifications. 16. IANA Considerations This document has no IANA actions. It does not define a new protocol, registry, media type, URI scheme, protocol parameter, or registered EAT Claim. Elekula Expires 20 May 2027 [Page 30] Internet-Draft AI Inference Attestation November 2026 17. Conclusion This document investigates two related but distinct attestation problems in AI inference environments. The first is establishing that an approved model artifact is the artifact actually admitted into an approved inference runtime. The second is establishing that an individual inference operation was processed by an execution instance whose relevant measurements have been appraised. The document identifies a candidate approach based on a protected model-loading path, trusted measurement collection, cryptographically protected Evidence, and a controlled request-dispatch mechanism where per-inference assurance is required. It distinguishes these mechanisms from ordinary application-level assertions and from freshness mechanisms that do not independently establish execution identity. Existing RATS specifications provide substantial foundations for representing component measurements and appraising Evidence. The remaining technical question is whether those specifications provide sufficient interoperable semantics for model-loading relationships and per-inference execution association. The proposal does not assume that a new protocol or Claim vocabulary is necessary. Further work should first evaluate existing EAT and measured-component mechanisms, identify the measurement capabilities required, and demonstrate the proposed binding through an implementation. The ultimate objective is interoperable Evidence that allows a Relying Party to distinguish an independently measured AI execution environment from an unverified assertion about which model an AI service claims to be running. 18. References 18.1. Normative References [RFC2119] Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", BCP 14, RFC 2119, March 1997. [RFC8174] Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC 2119 Key Words", BCP 14, RFC 8174, May 2017. Elekula Expires 20 May 2027 [Page 31] Internet-Draft AI Inference Attestation November 2026 [RFC9334] Birkholz, H., Thaler, D., Richardson, M., Smith, N., and W. Pan, "Remote ATtestation procedureS (RATS) Architecture", RFC 9334, January 2023. [RFC9711] Lundblade, L., Mandyam, G., O'Donoghue, J., and C. Wallace, "The Entity Attestation Token (EAT)", RFC 9711, April 2025. [RFC10013] Frost, S., Fossati, T., Tschofenig, H., and H. Birkholz, "Entity Attestation Token (EAT) Measured Component", RFC 10013, July 2026. 18.2. Informative References [PTV-AGENT] Damodaran, A., "The Prove-Transform-Verify (PTV) Protocol for Attested Agent Identity", draft-anandakrishnan-rats-ptv-agent-identity-01, September 2026, Work in Progress. [AEP-RATS] Sokolov, A., "Composing Application-Layer Action Evidence with Remote Attestation Procedures", draft-sokolov-rats-aep-composition-06, August 2026, Work in Progress. [RFC9943] Birkholz, H., et al., "Supply Chain Integrity, Transparency, and Trust (SCITT) Architecture", RFC 9943, 2025. [RATS-CHARTER] IETF, "Remote ATtestation ProcedureS (RATS) Working Group Charter", IETF Datatracker. Author's Address Kayode Oluwole Elekula Lagos Nigeria Email: kayodeelekula@gmail.com Elekula Expires 20 May 2027 [Page 32]