Skip to main content

Provenance-Attributed Inference Token (PAIT): A Protocol for Token-Level Inference Provenance and Identity-Conditioned Inference in Generative AI Systems
draft-vasylenko-pait-protocol-00

Document Type Active Internet-Draft (individual)
Author Oleh Vasylenko
Last updated 2026-05-16
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-vasylenko-pait-protocol-00
Internet Engineering Task Force                            O. Vasylenko
Internet-Draft                                    InterAlliance OU
Intended status: Standards Track                             16 May 2026
Expires: 17 November 2026

               Provenance-Attributed Inference Token (PAIT):
            A Protocol for Token-Level Inference Provenance and
            Identity-Conditioned Inference in Generative AI Systems
                 draft-vasylenko-pait-protocol-00

Abstract

   This document specifies the Provenance-Attributed Inference Token
   (PAIT) protocol, a wire-level protocol for token-level provenance
   attribution and identity-conditioned inference in Large Language
   Model (LLM) systems and other generative AI systems.  PAIT defines:
   (1) an Agent Identity Token format that binds a globally unique
   agent identifier to a hierarchical authorization level by means of
   an asymmetric digital signature; (2) a Provenance Manifest Record
   format that associates each output token of an inference session
   with verifiable attribution data referencing training-corpus
   segments; (3) a Trust Telemetry Signal format for aggregated session
   metrics emission to external monitoring endpoints; and (4) a wire
   protocol state machine governing the interaction between a
   requesting agent and a generative AI endpoint.  PAIT is intended to
   address the transparency obligations imposed by emerging regulatory
   frameworks for high-risk and general-purpose AI systems, in
   particular those concerning the provenance of AI-generated content
   at sub-document granularity.

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."

Vasylenko                Expires 17 November 2026               [Page 1]

Internet-Draft                    PAIT                       16 May 2026

   This Internet-Draft will expire on 17 November 2026.

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  . . . . . . . . . . . . . . . . . . . . . . . .   2
     1.1.  Motivation  . . . . . . . . . . . . . . . . . . . . . . .   3
     1.2.  Scope . . . . . . . . . . . . . . . . . . . . . . . . . .   4
     1.3.  Out of Scope  . . . . . . . . . . . . . . . . . . . . . .   4
   2.  Terminology . . . . . . . . . . . . . . . . . . . . . . . . .   4
     2.1.  Requirements Language . . . . . . . . . . . . . . . . . .   5
     2.2.  Definitions . . . . . . . . . . . . . . . . . . . . . . .   5
   3.  Protocol Overview . . . . . . . . . . . . . . . . . . . . . .   6
   4.  Agent Identity Token (PAIT-ID)  . . . . . . . . . . . . . . .   7
     4.1.  Token Structure . . . . . . . . . . . . . . . . . . . . .   7
     4.2.  Field Definitions . . . . . . . . . . . . . . . . . . . .   8
     4.3.  Authorization Levels  . . . . . . . . . . . . . . . . . .   9
     4.4.  Signature Algorithms  . . . . . . . . . . . . . . . . . .  10
     4.5.  Delegation  . . . . . . . . . . . . . . . . . . . . . . .  10
   5.  Provenance Manifest Record Format (PAIT-PM) . . . . . . . . .  11
     5.1.  Record Structure  . . . . . . . . . . . . . . . . . . . .  11
     5.2.  Per-Token Fields  . . . . . . . . . . . . . . . . . . . .  12
   6.  Trust Telemetry Signal Format (PAIT-TS) . . . . . . . . . . .  13
   7.  Wire Protocol State Machine . . . . . . . . . . . . . . . . .  14
   8.  Security Considerations . . . . . . . . . . . . . . . . . . .  15
   9.  IANA Considerations . . . . . . . . . . . . . . . . . . . . .  17
   10. Intellectual Property Rights  . . . . . . . . . . . . . . . .  18
   11. References  . . . . . . . . . . . . . . . . . . . . . . . . .  18
     11.1.  Normative References . . . . . . . . . . . . . . . . . .  18
     11.2.  Informative References . . . . . . . . . . . . . . . . .  19
   Author's Address  . . . . . . . . . . . . . . . . . . . . . . . .  20
   Appendix A.  Relationship to ITU-T Work Items . . . . . . . . . .  20

1.  Introduction

   Generative artificial intelligence systems, including Large Language
   Models (LLMs), increasingly produce content that is consumed by
   downstream applications, regulated parties, and end-users without

Vasylenko                Expires 17 November 2026               [Page 2]

Internet-Draft                    PAIT                       16 May 2026

   sub-document evidence of which sources influenced specific portions
   of the output.  Existing provenance frameworks for AI-generated
   content (e.g., [C2PA]) operate at the level of the complete media
   asset and certify origin of the asset as a whole.  General-purpose
   provenance models such as [W3C-PROV-DM] describe upstream data
   lineage but do not bind individual output units of an inference
   session to specific training-corpus segments at runtime.

   This document specifies the Provenance-Attributed Inference Token
   (PAIT) protocol, which addresses this gap at sub-document
   granularity.  PAIT associates each output token of an inference
   session with a verifiable attribution record referencing one or
   more training-corpus segments; conditions inference behavior on a
   cryptographically verified agent identity; and emits aggregated
   trust metrics to an external monitoring endpoint.

1.1.  Motivation

   Three concurrent developments motivate this work:

   *  Regulatory transparency obligations.  Article 50 of the
      Regulation (EU) 2024/1689 (EU AI Act) [EU-AI-ACT] requires
      providers and deployers of certain AI systems to ensure that
      outputs are marked as artificially generated or manipulated in a
      machine-readable format.  Recital 133 elaborates that such
      marking should support detection of AI-generated content.
      Operationalizing this obligation at sub-document granularity for
      sequential generative systems requires a per-token wire format
      that current standards do not provide.

   *  Identity-conditioned access control for AI agents.  As AI agents
      increasingly act on behalf of principals with differentiated
      authorization scopes, the inference behavior of generative AI
      endpoints must be conditioned on the verified identity and
      authorization level of the requesting agent, including selective
      access to portions of training corpora and to subsets of
      computational layers.  General-purpose authorization frameworks
      such as OAuth 2.0 [RFC6749] and its delegation extensions do not
      address inference-time corpus and layer selection.

   *  Verifiable per-token provenance.  Downstream consumers and
      auditors require machine-readable evidence linking specific
      output tokens to specific source-corpus segments, with
      cryptographic integrity guarantees.  Existing attribution
      mechanisms internal to LLM systems produce, at best, ad-hoc
      attribution traces that are not standardized and are not
      anchored in an immutable record.

Vasylenko                Expires 17 November 2026               [Page 3]

Internet-Draft                    PAIT                       16 May 2026

   PAIT is the wire-level binding layer that connects identity-
   conditioned inference, per-token provenance attribution, and
   tamper-evident provenance recording into a single protocol.

1.2.  Scope

   This document specifies:

   *  the wire format of the PAIT Agent Identity Token (Section 4);

   *  the wire format of the PAIT Provenance Manifest Record
      (Section 5);

   *  the wire format of the PAIT Trust Telemetry Signal (Section 6);

   *  the wire protocol state machine governing the exchange between a
      requesting agent and a generative AI endpoint that implements
      PAIT (Section 7);

   *  signature algorithm identifiers for use in PAIT-ID
      (Section 4.4);

   *  IANA-managed registries for PAIT identifiers (Section 9).

1.3.  Out of Scope

   The following are explicitly out of scope for this document:

   *  algorithmic methods for computing the attribution vector
      associated with each output token;

   *  numerical values, formulas, or computational procedures for
      determining trust coefficients of training-corpus segments;

   *  the internal architecture of the generative AI system that
      implements PAIT, including its training procedure, its
      computational layer structure, and its language-prioritization
      mechanism;

   *  storage-layer implementation of the immutable provenance
      registry beyond the requirement of an append-only cryptographic
      hash chain.

2.  Terminology

Vasylenko                Expires 17 November 2026               [Page 4]

Internet-Draft                    PAIT                       16 May 2026

2.1.  Requirements Language

   The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
   "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY",
   and "OPTIONAL" 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.

2.2.  Definitions

   Agent:  An entity that submits an inference request to a generative
      AI endpoint.  An Agent may be a human user, an autonomous AI
      system, or an automated service.

   Agent Identity Protocol:  The class of protocols defining a
      structure and format of an agent identifier comprising a
      globally unique machine-readable identifier, a hierarchical
      authorization level, a list of permitted operations, a list of
      prohibited operations, validity-period timestamps, a
      cryptographic signature over all preceding fields, and a
      protocol version field.  The Agent Identity Token specified in
      Section 4 of this document is a wire-format instance of an
      agent identity protocol.

   Agent Identity Token (PAIT-ID):  The wire-format identity object
      defined in Section 4 of this document, instantiating the agent
      identity protocol class.  Carries a globally unique agent
      identifier, an authorization level, a permitted-operations
      list, a validity period, and a cryptographic signature.

   Authorization Level:  A discrete value drawn from a finite ordered
      set indicating the scope of access granted to an Agent.  This
      document defines three baseline levels (L0, L1, L2) in
      Section 4.3.

   Delegating Agent:  An Agent of authorization level L0 that issues a
      PAIT-ID for an Agent of lower authorization level.

   Generative AI Endpoint:  The network-accessible service that
      receives inference requests, generates output sequences, and
      emits PAIT-PM and PAIT-TS records.

   Inference Session:  A single end-to-end interaction comprising one
      input query, one verified PAIT-ID, the resulting output
      sequence, and the corresponding PAIT-PM record.

   Provenance Manifest Record (PAIT-PM):  The per-session machine-
      readable document defined in Section 5.  Contains one entry per
      output token.

   Token:  A single discrete output unit produced by the generative AI
      endpoint.  The granularity of a Token is determined by the
      tokenizer of the endpoint and is opaque to PAIT.

   Trust Telemetry Signal (PAIT-TS):  The aggregated per-session
      metrics object defined in Section 6.  Transmitted to an external
      monitoring endpoint.

Vasylenko                Expires 17 November 2026               [Page 5]

Internet-Draft                    PAIT                       16 May 2026

   Validity Period:  The closed time interval, bounded by
      validity_start_utc and validity_end_utc, during which a PAIT-ID
      is considered valid for the purpose of cryptographic
      verification.

3.  Protocol Overview

   PAIT is a request-response protocol with four atomic wire-format
   objects:

           +---------------+         +---------------+
           |   Requesting  |         |  Generative   |
           |     Agent     |         |  AI Endpoint  |
           +-------+-------+         +-------+-------+
                   |                         |
                   |  Inference Request +    |
                   |  PAIT-ID               -|--> [Verify]
                   |                         |
                   |                         |--> [Route]
                   |                         |
                   |                         |--> [Generate]
                   |                         |
                   |  Output Sequence +      |
                   |  PAIT-PM Record  <------|
                   |                         |
                                             |
                                             v
                              +-----------------------------+
                              |  External Trust Monitoring  |
                              |          Endpoint           |
                              +-------------+---------------+
                                            ^
                                            |  PAIT-TS Signal
                                            |  (HTTPS REST)
                              +-------------+---------------+
                              |  Generative AI Endpoint     |
                              +-----------------------------+

                       Figure 1: PAIT Protocol Overview

   The four atomic objects are:

   *  PAIT-ID (Section 4): identifies and authorizes the requesting
      agent.

   *  PAIT-PM (Section 5): records per-token provenance for the output
      sequence.

Vasylenko                Expires 17 November 2026               [Page 6]

Internet-Draft                    PAIT                       16 May 2026

   *  PAIT-TS (Section 6): emits aggregated session metrics to an
      external monitoring endpoint.

   *  PAIT Registry Record: not directly transmitted on the wire as a
      protocol element, but referenced by PAIT-PM and PAIT-TS as the
      immutable cryptographic record of inference sessions and corpus
      segments.  Registry record format is out of scope for this
      document.

4.  Agent Identity Token (PAIT-ID)

4.1.  Token Structure

   A PAIT-ID is a JSON object with the following top-level fields:

   {
     "protocol_version": "1.0",
     "gai":              "<UUID>",
     "auth_level":       "<L0|L1|L2>",
     "allowed_ops":      [ "<op>", ... ],
     "prohibited_ops":   [ "<op>", ... ],
     "validity_start_utc":"<ISO 8601 timestamp>",
     "validity_end_utc":  "<ISO 8601 timestamp>",
     "signature": {
       "alg":  "<ES256|EdDSA>",
       "kid":  "<key identifier>",
       "value":"<base64url-encoded signature>"
     }
   }

                    Figure 2: PAIT-ID JSON structure

   The signature value is computed over the canonicalized JSON
   serialization of all preceding fields (protocol_version through
   validity_end_utc) using the algorithm identified in signature.alg.
   The canonicalization procedure is specified by [RFC8785].

4.2.  Field Definitions

   protocol_version:
      A string identifying the PAIT protocol version.  This document
      specifies version "1.0".  Implementations MUST reject PAIT-IDs
      with an unrecognized protocol_version.

   gai (Global Agent Identifier):
      A globally unique machine-readable identifier in UUID version 4
      format [RFC4122].  Each Agent SHOULD be assigned exactly one
      gai value across all PAIT deployments.

Vasylenko                Expires 17 November 2026               [Page 7]

Internet-Draft                    PAIT                       16 May 2026

   auth_level:
      A string indicating the hierarchical authorization level
      assigned to the Agent.  This document defines three baseline
      values ("L0", "L1", "L2") in Section 4.3.  Extensions to this
      document MAY define additional values.

   allowed_ops:
      A JSON array of machine-readable operation identifiers that the
      Agent is permitted to invoke against the Generative AI
      Endpoint.  An empty array signifies that all operations
      permissible at the assigned auth_level are allowed.  Operation
      identifiers are deployment-specific strings.

   prohibited_ops:
      A JSON array of machine-readable operation identifiers that the
      Agent is explicitly forbidden from invoking, irrespective of
      auth_level.  An empty array signifies no explicit prohibitions.
      Where allowed_ops and prohibited_ops both contain the same
      identifier, prohibited_ops takes precedence.

   validity_start_utc, validity_end_utc:
      ISO 8601 [ISO8601] timestamps in UTC defining the closed time
      interval during which the PAIT-ID is valid.  Implementations
      MUST reject PAIT-IDs whose current verification time falls
      outside this interval and MUST assign the minimum
      authorization level (defined as the lowest-privilege level in
      the implementation's level set, typically L2) to the requesting
      Agent.

   signature.alg:
      An identifier for the digital signature algorithm used.  This
      document defines two values: "ES256" (ECDSA on curve P-256, per
      [RFC7518] Section 3.4) and "EdDSA" (Edwards-curve DSA, per
      [RFC8037]).

   signature.kid:
      An opaque key identifier referencing the public key against
      which signature.value is verified.  The mapping from kid to
      public key material is out of scope for this document.

   signature.value:
      The base64url-encoded signature, computed as described in
      Section 4.1.

4.3.  Authorization Levels

   This document defines three baseline authorization levels:

   *  L0 (full):  Highest privilege.  Grants the requesting Agent

Vasylenko                Expires 17 November 2026               [Page 8]

Internet-Draft                    PAIT                       16 May 2026

      access to the full training corpus available to the Generative
      AI Endpoint and to all computational layers of the underlying
      model.  L0 Agents MAY act as Delegating Agents (Section 4.5).

   *  L1 (verified):  Intermediate privilege.  Grants the requesting
      Agent access to a subset of the training corpus restricted to
      segments bearing open licenses (e.g., CC0, CC BY 4.0,
      Apache 2.0) and to a subset of the computational layers of the
      underlying model.

   *  L2 (public):  Minimum privilege.  Grants the requesting Agent
      access to the public subset of the training corpus (e.g.,
      segments bearing CC0 or Public Domain) and to a restricted
      subset of computational layers.

   The mapping from auth_level to specific corpus subset and to
   specific computational-layer subset is implementation-specific.
   PAIT requires only that the mapping be deterministic and that L0 ⊇
   L1 ⊇ L2 with respect to accessible corpus and layers.

4.4.  Signature Algorithms

   Implementations MUST support at least one of "ES256" and "EdDSA".
   Implementations SHOULD support both.  Additional algorithms MAY be
   registered via the IANA registry defined in Section 9.

4.5.  Delegation

   An Agent of authorization level L0 MAY issue PAIT-IDs for Agents
   of lower authorization level (L1 or L2).  A delegated PAIT-ID:

   *  MUST be signed by the Delegating Agent;

   *  MUST contain an allowed_ops value that is a subset of the
      Delegating Agent's own allowed_ops;

   *  MUST have a validity_end_utc not later than the validity_end_utc
      of the Delegating Agent's PAIT-ID;

   *  inherits access only to the corpus subset and layer subset of
      its own (lower) auth_level, regardless of the corpus and layer
      subset accessible to the Delegating Agent.

5.  Provenance Manifest Record Format (PAIT-PM)

5.1.  Record Structure

   A PAIT-PM record is emitted by the Generative AI Endpoint for each

Vasylenko                Expires 17 November 2026               [Page 9]

Internet-Draft                    PAIT                       16 May 2026

   Inference Session and is transmitted to the requesting Agent
   together with the output sequence.  A PAIT-PM record is encoded as
   a JSON Lines [JSONL] document in which each line corresponds to
   exactly one output token of the Inference Session.

   The first line of a PAIT-PM record MUST be a header object with
   the following fields:

   {
     "type":            "pait-pm-header",
     "protocol_version":"1.0",
     "session_id":      "<UUID>",
     "agent_id":        "<UUID>",  // gai from PAIT-ID
     "start_utc":       "<ISO 8601 timestamp>",
     "model_id":        "<endpoint-specific identifier>",
     "prev_session_hash":"<base64url-encoded SHA-256 digest>"
   }

                  Figure 3: PAIT-PM header object structure

   Each subsequent line is a per-token record (Section 5.2).

5.2.  Per-Token Fields

   {
     "type":            "pait-pm-token",
     "token_idx":       <integer>,
     "token_repr":      "<opaque endpoint-specific representation>",
     "attribution":     [
       {
         "segment_id":  "<segment identifier>",
         "weight":      <number in [0,1]>,
         "license":     "<license class string>"
       },
       ...
     ],
     "license_purity":  <number in [0,1]>
   }

                Figure 4: PAIT-PM per-token record structure

   Field semantics:

   token_idx:
      Zero-based index of the token within the output sequence.

   token_repr:
      An opaque representation of the output token in the
      endpoint-specific tokenizer.  PAIT does not constrain this

Vasylenko                Expires 17 November 2026              [Page 10]

Internet-Draft                    PAIT                       16 May 2026

      representation.

   attribution:
      A JSON array of attribution entries.  Each entry references a
      training-corpus segment that contributed to the generation of
      this token.  The weight field denotes the relative contribution
      of the referenced segment.  The sum of weight values across all
      entries SHOULD be 1.0 (within rounding tolerance), unless the
      attribution vector has been renormalized after filtering.
      The license field denotes the license class associated with the
      referenced segment.

   license_purity:
      A scalar in the closed interval [0,1] denoting an aggregate
      indicator of license-compatible attribution for this token.
      The method by which license_purity is computed is out of scope
      for this document.

   The final line of a PAIT-PM record MUST be a footer object:

   {
     "type":             "pait-pm-footer",
     "session_id":       "<UUID>",
     "token_count":      <integer>,
     "end_utc":          "<ISO 8601 timestamp>",
     "manifest_hash":    "<base64url-encoded SHA-256 digest>"
   }

                Figure 5: PAIT-PM footer object structure

   The manifest_hash is computed by an implementation-defined
   deterministic function over the canonical representation of the
   header object and all per-token records, in order.  The hash
   function MUST be SHA-256 [RFC6234] or stronger.  Implementations
   MUST ensure identical output for identical input.

6.  Trust Telemetry Signal Format (PAIT-TS)

   A PAIT-TS signal is emitted by the Generative AI Endpoint to an
   external monitoring endpoint upon completion of an Inference
   Session, or upon detection of an anomalous condition during a
   session.

   A PAIT-TS signal is a JSON object transmitted over HTTPS to a
   pre-configured REST endpoint:

   {
     "type":               "pait-ts-signal",
     "protocol_version":   "1.0",
     "session_id":         "<UUID>",

Vasylenko                Expires 17 November 2026              [Page 11]

Internet-Draft                    PAIT                       16 May 2026

     "agent_id":           "<UUID>",
     "auth_level":         "<L0|L1|L2>",
     "token_count":        <integer>,
     "license_distribution":{
       "<license class>":  <number in [0,1]>,
       ...
     },
     "high_trust_fraction":<number in [0,1]>,
     "timestamp_utc":      "<ISO 8601 timestamp>",
     "signal_mode":        "<batch|out_of_order>",
     "anomaly_flag":       <boolean>,
     "endpoint_signature": {
       "alg":   "<ES256|EdDSA>",
       "kid":   "<key identifier>",
       "value": "<base64url-encoded signature>"
     }
   }

                  Figure 6: PAIT-TS signal structure

   Implementations MUST transmit PAIT-TS signals over a confidential
   and integrity-protected channel.  TLS 1.3 [RFC8446] is REQUIRED.

   Implementations MUST support two transmission modes:

   *  batch: aggregated signals transmitted at a configured interval,
      not less frequent than once per 24 hours;

   *  out_of_order: immediate transmission upon detection of an
      anomalous condition.  The conditions classified as anomalous
      and the corresponding thresholds are out of scope for this
      document.  When signal_mode is "out_of_order", anomaly_flag
      MUST be set to true.

7.  Wire Protocol State Machine

   The PAIT protocol state machine governs the interaction between
   the requesting Agent and the Generative AI Endpoint.  Nine states
   are defined:

   S0 (Idle):
      The Endpoint awaits an inference request.

   S1 (Request Received):
      The Endpoint has received an inference request containing a
      PAIT-ID.  Transitions to S2.

   S2 (Identity Verifying):

Vasylenko                Expires 17 November 2026              [Page 12]

Internet-Draft                    PAIT                       16 May 2026

      The Endpoint verifies the PAIT-ID:

      *  validates protocol_version;

      *  validates that the current verification time is within
         [validity_start_utc, validity_end_utc];

      *  retrieves the public key identified by signature.kid;

      *  verifies signature.value against the canonicalized
         serialization of preceding fields using signature.alg.

      On verification failure (any of: absent PAIT-ID, invalid
      signature, expired validity period, unrecognized
      protocol_version), the Endpoint MUST assign the minimum
      authorization level (typically L2) to the requesting Agent and
      transition to S3.  On verification success, the Endpoint
      transitions to S3 with the verified auth_level.

   S3 (Inference Routing):
      The Endpoint selects:

      *  the subset of the training corpus accessible at the verified
         auth_level;

      *  the subset of computational layers accessible at the verified
         auth_level;

      *  the processing mode for the inference request.

      The Endpoint transitions to S4.

   S4 (Generation):
      The Endpoint generates the output sequence.  For each output
      token, the Endpoint computes attribution data as defined by the
      Endpoint's internal attribution method (out of scope for this
      document) and appends a per-token PAIT-PM record to the session
      manifest.  The Endpoint transitions to S5 after the final token
      of the sequence has been generated.

   S5 (Manifest Sealing):
      The Endpoint emits the PAIT-PM footer (Section 5.2), computes
      the manifest_hash over the entire manifest, and seals the
      manifest.  The Endpoint transitions to S6.

   S6 (Registry Append):
      The Endpoint writes a new record of session-completion type to
      its immutable cryptographic provenance registry.  The new

Vasylenko                Expires 17 November 2026              [Page 13]

Internet-Draft                    PAIT                       16 May 2026

      record incorporates the SHA-256 hash of the immediately
      preceding registry record into its prev_hash field.  The
      Endpoint transitions to S7.

   S7 (Response Transmission):
      The Endpoint transmits the output sequence and the sealed
      PAIT-PM record to the requesting Agent.  The Endpoint
      transitions to S8.

   S8 (Telemetry Emission):
      The Endpoint emits a PAIT-TS signal (Section 6) to the
      configured external monitoring endpoint.  If an anomalous
      condition was detected during S4, the Endpoint MUST emit the
      PAIT-TS signal in "out_of_order" mode with anomaly_flag set to
      true; otherwise, the Endpoint MAY defer emission to a
      subsequent batch.  After emission (or deferral), the Endpoint
      transitions back to S0.

   The state machine guarantees that no PAIT-PM record may be
   transmitted to a requesting Agent without prior identity
   verification (S2) and prior registry append (S6).

8.  Security Considerations

   This section addresses the principal security considerations
   applicable to PAIT.

   Identity verification.  The integrity of the PAIT protocol
   depends critically on the cryptographic verification of PAIT-IDs
   in state S2.  Implementations MUST:

   *  reject PAIT-IDs with invalid signatures;

   *  reject PAIT-IDs whose verification time falls outside the
      [validity_start_utc, validity_end_utc] interval;

   *  assign the minimum authorization level in any case of
      verification failure.

   Replay protection.  PAIT-IDs are time-bounded by the validity
   period.  Implementations SHOULD additionally maintain a record
   of recently observed PAIT-IDs (identified by signature.value) and
   reject duplicates received within a configurable window, to
   defend against replay attacks within an otherwise valid
   validity period.

   Confidentiality.  PAIT-ID, PAIT-PM, and PAIT-TS objects may
   contain information that the requesting Agent or the operator

Vasylenko                Expires 17 November 2026              [Page 14]

Internet-Draft                    PAIT                       16 May 2026

   considers sensitive.  Implementations MUST transport PAIT objects
   over a confidentiality- and integrity-protected channel.  TLS 1.3
   [RFC8446] is REQUIRED.

   Key management.  The mapping from signature.kid to public key
   material is out of scope for this document.  Implementations
   SHOULD use established key-distribution mechanisms such as JWKS
   [RFC7517] or DNS-based key discovery.

   Registry integrity.  The cryptographic hash chain of the
   immutable provenance registry provides tamper-evident integrity:
   modification of any registry record causes a mismatch of the
   record_hash and of prev_hash in all subsequent records.
   Implementations MUST store the registry in append-only mode.
   Implementations SHOULD provide a read-only audit interface for
   external verification.

   Trust telemetry endpoint authentication.  The Endpoint emitting
   PAIT-TS signals MUST authenticate the external monitoring
   endpoint to which signals are transmitted, e.g., via mutual TLS,
   to prevent diversion of telemetry to an attacker-controlled
   endpoint.  PAIT-TS signals MUST be signed by the Endpoint using
   endpoint_signature (Section 6).

   Privacy.  Implementations MUST consider that token_idx,
   attribution, and timestamp fields collectively may permit
   inference of user query content.  Implementations SHOULD
   evaluate compatibility with applicable privacy frameworks
   (e.g., [GDPR]) prior to deployment.

   Compatibility with self-prior-art exclusions.  This protocol
   specifies wire formats and state-machine behavior only.
   Algorithmic methods for attribution-vector computation, trust-
   coefficient determination, and routing-decision logic are
   intentionally out of scope, as described in Section 1.3.

9.  IANA Considerations

   This document requests IANA to create three new registries under
   the "PAIT Parameters" registry group:

   PAIT-ID Signature Algorithms registry:
      Initial values:
      *  "ES256" (ECDSA on curve P-256, per [RFC7518] Section 3.4)
      *  "EdDSA" (Edwards-curve DSA, per [RFC8037])
      Registration policy: Specification Required [RFC8126].

   PAIT-ID Authorization Levels registry:

Vasylenko                Expires 17 November 2026              [Page 15]

Internet-Draft                    PAIT                       16 May 2026

      Initial values:
      *  "L0" (full)
      *  "L1" (verified)
      *  "L2" (public)
      Registration policy: Specification Required [RFC8126].

   PAIT Protocol Versions registry:
      Initial value:
      *  "1.0" (this document)
      Registration policy: Standards Action [RFC8126].

10. Intellectual Property Rights

   The technology described in this document is the subject of an
   international patent application, PCT/IB2026/053131 (priority date
   25 March 2026, 19 claims, International Searching Authority:
   European Patent Office).  The patent holder is willing to grant
   a license for any standards-essential patent claims under fair,
   reasonable, and non-discriminatory (FRAND) terms with reciprocity.
   A formal IPR disclosure has been filed with the IETF.  See the
   IETF IPR disclosure record for the current declaration.

11. References

11.1.  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/info/rfc2119>.

   [RFC4122]    Leach, P., Mealling, M., and R. Salz, "A Universally
                Unique IDentifier (UUID) URN Namespace", RFC 4122,
                July 2005,
                <https://www.rfc-editor.org/info/rfc4122>.

   [RFC6234]    Eastlake 3rd, D. and T. Hansen, "US Secure Hash
                Algorithms (SHA and SHA-based HMAC and HKDF)",
                RFC 6234, May 2011,
                <https://www.rfc-editor.org/info/rfc6234>.

   [RFC7517]    Jones, M., "JSON Web Key (JWK)", RFC 7517, May 2015,
                <https://www.rfc-editor.org/info/rfc7517>.

   [RFC7518]    Jones, M., "JSON Web Algorithms (JWA)", RFC 7518,

Vasylenko                Expires 17 November 2026              [Page 16]

Internet-Draft                    PAIT                       16 May 2026

                May 2015,
                <https://www.rfc-editor.org/info/rfc7518>.

   [RFC8037]    Liusvaara, I., "CFRG Elliptic Curve Diffie-Hellman
                (ECDH) and Signatures in JSON Object Signing and
                Encryption (JOSE)", RFC 8037, January 2017,
                <https://www.rfc-editor.org/info/rfc8037>.

   [RFC8126]    Cotton, M., Leiba, B., and T. Narten, "Guidelines for
                Writing an IANA Considerations Section in RFCs",
                BCP 26, RFC 8126, June 2017,
                <https://www.rfc-editor.org/info/rfc8126>.

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

   [RFC8446]    Rescorla, E., "The Transport Layer Security (TLS)
                Protocol Version 1.3", RFC 8446, August 2018,
                <https://www.rfc-editor.org/info/rfc8446>.

   [RFC8785]    Rundgren, A., Jordan, B., and S. Erdtman, "JSON
                Canonicalization Scheme (JCS)", RFC 8785, June 2020,
                <https://www.rfc-editor.org/info/rfc8785>.

   [ISO8601]    International Organization for Standardization,
                "Date and time -- Representations for information
                interchange -- Part 1: Basic rules",
                ISO 8601-1:2019, February 2019.

11.2.  Informative References

   [C2PA]       Coalition for Content Provenance and Authenticity,
                "C2PA Technical Specification, Version 2.2",
                Joint Development Foundation, May 2025,
                <https://c2pa.org/specifications/specifications/2.2/
                specs/C2PA_Specification.html>.

   [W3C-PROV-DM]
                Moreau, L. and P. Missier, Eds., "PROV-DM: The PROV
                Data Model", W3C Recommendation, 30 April 2013,
                <https://www.w3.org/TR/prov-dm/>.

   [EU-AI-ACT]  European Parliament and Council, "Regulation (EU)
                2024/1689 of the European Parliament and of the
                Council of 13 June 2024 laying down harmonised rules
                on artificial intelligence (Artificial Intelligence
                Act)", Official Journal of the European Union,

Vasylenko                Expires 17 November 2026              [Page 17]

Internet-Draft                    PAIT                       16 May 2026

                July 2024,
                <https://eur-lex.europa.eu/eli/reg/2024/1689/oj>.

   [GDPR]       European Parliament and Council, "Regulation (EU)
                2016/679 of the European Parliament and of the
                Council of 27 April 2016 on the protection of
                natural persons with regard to the processing of
                personal data (General Data Protection Regulation)",
                Official Journal of the European Union, April 2016,
                <https://eur-lex.europa.eu/eli/reg/2016/679/oj>.

   [RFC6749]    Hardt, D., Ed., "The OAuth 2.0 Authorization
                Framework", RFC 6749, October 2012,
                <https://www.rfc-editor.org/info/rfc6749>.

   [JSONL]      "JSON Lines", <https://jsonlines.org/>.

   [ITU-T-X-LLMP-TIP]
                Vasylenko, O., "Proposed new work item X.llmp-tip:
                LLM token-level inference provenance", ITU-T SG17
                Contribution T25-SG17-C-0461, 15 May 2026,
                <https://www.itu.int/md/T25-SG17-C-0461/>.

Appendix A.  Relationship to ITU-T Work Items

   This protocol specification is a companion document to ITU-T
   SG17 Contribution C-0461 (X.llmp-tip), which establishes the
   requirements framework for token-level LLM provenance within
   the ITU-T standardization process.

   The requirements in X.llmp-tip operate at the conceptual level.
   This IETF Internet-Draft provides a protocol-level realization,
   defining wire formats, message structures, and state transitions.

   The two documents are intended for parallel development:
   X.llmp-tip within ITU-T SG17 (security and trust requirements)
   and this draft within the IETF (protocol specification).

   Related ITU-T contributions from the same contributor:

      C-0456 (X.gaiv):  Globally interoperable agent identity
         verification.

      C-0457 (X.aaid):  Agent authentication and identity
         delegation.

      C-0458 (X.rcae):  Regulatory continuous attestation engine
         for AI agents.

      C-0459 (X.atcp):  Agent trust confidence primitives.

      C-0460 (X.aam):  AI accountability manifest.

   These six contributions together define the Agentic Trust
   Control Plane architecture within ITU-T SG17.

Author's Address

   Oleh Vasylenko
   InterAlliance OU
   Tallinn
   Estonia

   Email: oleg@vasylenko.tel

Vasylenko                Expires 17 November 2026              [Page 18]