Asqav

Portugal joaoagm90@gmail.com

General AI Agents Audit Trail Compliance EU AI Act DORA NIST AI RMF Colorado AI Act HIPAA NYDFS CIRCIA This document defines a multi-jurisdiction compliance profile of the signed action receipt format used by AI agents to record machine-readable evidence of access-control decisions. The profile binds receipt fields to two regulatory surfaces: the European Union obligations of Articles 12 and 26 of Regulation (EU) 2024/1689 (the EU AI Act) and Article 17 of Regulation (EU) 2022/2554 (DORA), and the United States obligations of the NIST Artificial Intelligence Risk Management Framework, the Colorado Artificial Intelligence Act (SB 24-205), the Texas Responsible AI Governance Act (HB 149), the New York Department of Financial Services Cybersecurity Regulation (23 NYCRR Part 500), the HIPAA Security Rule (45 CFR Part 164, Subpart C), SEC Rule 17a-4 (17 CFR 240.17a-4), and the Cyber Incident Reporting for Critical Infrastructure Act of 2022 (CIRCIA). It does not redefine the wire format, the canonicalization transformation, or the signing algorithms of the underlying receipt format; it narrows the set of values to which the canonicalization rule is applied (IEEE-754 floating-point numbers MUST NOT appear in digest-covered fields) and tightens a small number of pre-existing wire surfaces under this revision (see ). It tightens a subset of the OPTIONAL fields to REQUIRED, imposes a retention floor, requires at least one timestamping anchor (RFC 3161 or OpenTimestamps; both RECOMMENDED), and adds three extension fields. The third, counterparty_binding, is added in this revision to provide cryptographic cross-agent byte-equality evidence under a compromised intermediary.

Introduction

Profile, Not Fork specifies a generic, signed receipt envelope for recording machine-to-machine access control decisions made by AI agents. Section 2.2 of defines a common payload field set in which all fields except type, issued_at, and issuer_id are OPTIONAL. Section 5.7 of introduces hash chaining (previousReceiptHash) inside an optional Commitment Mode extension. does not define receipt retention, does not require timestamping anchors, and does not bind to any regulatory regime. This document is an additive overlay on : it constrains fields the upstream draft leaves OPTIONAL, fixes their values where regulation requires, and adds three extension fields with reserved names (risk_class, incident_class, and counterparty_binding). A Compliance Receipt remains a conformant receipt. Field references use upstream field names rather than section numbers, to reduce maintenance hazard if upstream re-numbers in a future revision.

Scope This document fills the regulatory binding gap on two surfaces. Section 5 binds the receipt to European Union obligations: Article 12 (record-keeping) and Article 26 (deployer obligations) of the EU AI Act, and Article 17 (ICT-related incident management) of DORA. Section 6 binds the receipt to United States obligations: the voluntary functions of the NIST AI Risk Management Framework, the deployer obligations of the Colorado AI Act and the Texas Responsible AI Governance Act, the audit-trail and incident-reporting obligations of NYDFS Part 500, the audit controls and documentation retention of the HIPAA Security Rule, the broker-dealer recordkeeping requirements of SEC Rule 17a-4, and the covered-incident reporting requirements of CIRCIA. The bindings are written from the Deployer's perspective, where Deployer is used in the regime-specific sense (Article 3(4) of for EU bindings; Section 6-1-1701(6) of the Colorado Revised Statutes for Colorado bindings). Where another statute uses a different term (Provider, Financial Entity, Covered Entity for HIPAA, Covered Entity for NYDFS, Broker-Dealer for SEC, Covered Entity for CIRCIA), the binding section names the term as the source statute uses it. A verifier that implements only can cryptographically validate a profile receipt but cannot attest the additional compliance bindings of this document.

Conventions and Definitions 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 when, and only when, they appear in all capitals, as shown here. The following terms are used in this document.

Action:

An operation performed by an AI agent that is subject to a policy evaluation. Examples include a tool invocation, an external API call, a write to durable storage, and the issuance of an irreversible instruction to another system.

Action Receipt:

A signed envelope conforming to that records the policy evaluation result for a single Action.

Compliance Receipt:

An Action Receipt that additionally satisfies the requirements of this profile.

Deployer (EU AI Act):

As defined in Article 3(4) of .

Deployer (Colorado AI Act):

As defined in Section 6-1-1701(6) of the Colorado Revised Statutes, as enacted by .

High-Risk AI System (EU AI Act):

As defined in Article 6 of .

High-Risk AI System (Colorado AI Act):

As defined in Section 6-1-1701(9) of the Colorado Revised Statutes, as enacted by .

Financial Entity:

As defined in Article 2(2) of , for entities listed in Article 2(1).

Covered Entity (HIPAA):

As defined in 45 CFR 160.103, namely a health plan, a health care clearinghouse, or a health care provider that transmits health information in electronic form in connection with a covered transaction.

Covered Entity (NYDFS):

As defined in 23 NYCRR 500.1(e), namely any person operating under or required to operate under a license, registration, charter, certificate, permit, accreditation or similar authorization under the Banking Law, the Insurance Law or the Financial Services Law, regardless of whether the covered entity is also regulated by other government agencies.

Broker-Dealer:

As defined in section 3(a)(4) and 3(a)(5) of the Securities Exchange Act of 1934, subject to recordkeeping under .

Covered Entity (CIRCIA):

As to be defined in the final rule promulgated under the Cyber Incident Reporting for Critical Infrastructure Act of 2022. Pending publication of the final rule, the term is interpreted in accordance with the statutory definition at 6 U.S.C. 681 and CISA's notice of proposed rulemaking.

Audit Pack:

A bundle of Compliance Receipts, the chain commitments that link them, the public verification keys, the trust anchor metadata, and the regime mapping required by Sections 5 and 6 of this document, packaged for delivery to a regulator or auditor.

Relationship to ACTA-RECEIPTS This profile is an additive overlay on . It does not modify the envelope, the canonicalization rule, the signature object, or the algorithm registry of . The following normative statements apply.

• Implementations of this profile MUST produce receipts that are cryptographically verifiable by a conformant verifier under the canonicalization rules (JCS, ) and the signature scope of Section 5.6.
• Implementations of this profile MUST NOT introduce new top-level fields in the signed payload that conflict with names reserved by .
• Implementations of this profile MAY use any signature algorithm permitted by : EdDSA (Ed25519, mandatory-to-implement, ), ES256 (ECDSA using P-256 and SHA-256, ), and ML-DSA-65 ().
• Where marks a field OPTIONAL and this profile marks the same field REQUIRED, the stricter requirement applies to Compliance Receipts.

A receipt that fails any MUST clause of this profile is not a Compliance Receipt. It MAY still be a valid receipt. This profile differentiates from on three axes: mandatory hash-chain linkage (upstream Commitment Mode is OPTIONAL), mandatory anchoring with RFC 3161 or OpenTimestamps (both RECOMMENDED; upstream lists Sigstore Rekor in its Implementation Status appendix as an OPTIONAL temporal anchor), and a retention floor tied to specific regulatory articles (upstream is silent on retention).

Canonicalization Scope This section is normative. The canonicalization rule itself (JCS, ) is inherited unchanged from ; this section bounds the inputs the rule is applied to, so that the cross-implementation byte equality on which the hash chain of , the anchor scope of , and the cross-agent binding of all depend is achievable in practice. IEEE-754 floating-point numbers MUST NOT appear in the canonical form covered by a SHA-256 digest under this profile. Callers MUST serialize numeric values that are not exact integers in the IEEE-754 safe integer range (the closed interval from minus (2 to the 53 minus 1) to plus (2 to the 53 minus 1) inclusive) either as JSON strings or as integer-rational pairs (numerator and denominator as JSON numbers within that safe integer range) before the canonicalization step. Rationale: Section 3.2.2.3 of specifies, by reference to Section 7.1.12.1 of ECMA-262, a byte-stable serialization that in principle covers all IEEE-754 double-precision values, integer or fractional. In practice, several widely deployed JSON serializers do not implement the ECMA-262 Number-to-String algorithm with byte fidelity: Python json.dumps, Go encoding/json, and Java Jackson (without explicit configuration) all produce different byte sequences from the same IEEE-754 double in documented cases (round-to-even ties, subnormal values, large-magnitude values requiring scientific notation). Compliance and regulatory contexts (monetary amounts, retention thresholds, anchoring intervals) additionally prefer exact integer or string-encoded decimal representations because float rounding loses the exact bytes that auditors quote. This profile therefore shifts the canonicalization burden off implementers (who would otherwise have to verify ECMA-262 conformance of an underlying JSON library) and onto callers, who are in a better position to choose a portable representation for the use case at hand. A receipt that carries a floating-point number in a digest-covered field is not guaranteed to verify across implementations even when each implementation independently conforms to , because the conformance burden has not been met by every mainstream JSON library. Tool-version-specific semantic equivalence is OUT OF SCOPE for the chain layer of this profile. The chain layer guarantees byte equality only. Examples of semantic equivalence that this profile does not assert and does not require a verifier to assert: SQL keyword case folding (SELECT vs select), filesystem path normalization (trailing slash, redundant separators, symlink resolution), Unicode normalization in any form (NFC, NFD, NFKC, NFKD); Section 3.1 of requires that all components depending on JCS preserve Unicode string data as-is, and Section 3.2.2.2 of serializes each code point without normalization, so callers MUST NOT rely on a verifier normalizing strings before comparison, locale-aware string collation (Turkish dotted-i, German sharp-s case folding, ICU collation tables), numeric tolerance (1.0 vs 1, 1e3 vs 1000), or URL percent-encoding choices below the RFC 3986 unreserved set. Higher-level semantic equivalence is a per-tool concern and, where required by a regulator, MUST be expressed in the policy artefact resolved through policy_digest () rather than in the chain. The chain layer of this profile answers a single question for a verifier or a regulator: did the same canonicalized bytes pass through agent X at wall-clock time T, as fixed by the anchor evidence of . Anything beyond that question, including whether two byte sequences are semantically equivalent under a downstream tool, whether a policy update materially changed the meaning of a previously accepted Action, or whether a counterparty's interpretation of the same bytes matched the originator's, is the verifier's concern and is supported by the Audit Pack manifest () and the verifier reporting fields of , not by the chain itself.

Receipt Field Profile This section enumerates fields defined by and states the additional requirements that this profile places on them. Field names follow exactly. Compliance Receipts MUST use the upstream wire field name signature for the signature object, exactly as defined in Sections 2.1 and 2.1.1 of . The keys inside that object are alg, kid, sig. Implementations whose internal storage uses a different field name MUST translate to signature on emission and on canonicalization for verification; receipts that appear on the wire under any other top-level field name are non-conformant to and to this profile. Anchors MUST be projected into a top-level anchors array with a type discriminator and a value field carrying the anchor bytes (base64-encoded for binary payloads). Flat-column implementations MUST project on emission and Audit Pack export.

Common Payload Fields

type Compliance Receipts MUST set type to a value drawn from the namespace protectmcp:decision, protectmcp:restraint, or protectmcp:lifecycle, or to an extension namespace registered for use with this profile.

issued_at REQUIRED upstream and in this profile. The value MUST be an ISO 8601 timestamp with an explicit timezone. The producing system MUST source the value from a clock synchronized to a recognized time authority and MUST NOT backdate the value. Verifiers MUST reject receipts whose issued_at is more than 300 seconds ahead of the verifier's own clock. Verifiers MUST NOT reject a receipt solely because issued_at lies in the past; past skew is bounded by the applicable retention floor in Sections 5 and 6, not by freshness. Historical receipts within retention MUST verify on the same path as fresh ones.

issuer_id REQUIRED upstream and in this profile. The value MUST identify a legal entity, not a natural person. Where the producing system is operated by a Deployer, the issuer_id MUST resolve, through the trust anchor metadata in the Audit Pack, to a record naming the Deployer. To preserve the upstream Section 2.2 invariant that issuer_id MUST match the kid field of the signature object, Compliance Receipts MUST place the same value in both issuer_id and kid; the verifier resolves that value to a public key through the Audit Pack trust-anchor metadata rather than through the well-known JWK Set endpoint or the RECOMMENDED sb:issuer:<base58-fingerprint> form of Section 2.1.1. This profile thereby supersedes the upstream RECOMMENDED kid format for Compliance Receipts; the upstream RECOMMENDED format remains valid for non-Compliance receipts. Implementations SHOULD use a Legal Entity Identifier (LEI) as defined by where one is allocated to the Deployer. Examples and test fixtures MUST use a placeholder whose four-character LOU prefix (positions 1-4) is not allocated in the GLEIF Local Operating Unit code list, whose positions 5-6 are the ISO 17442 reserved value 00, and whose two trailing characters (positions 19-20) are the ISO 7064 mod 97-10 check digits computed over positions 1-18 (for example 00000000000000000098, where the all-zero 18-character base produces the check digits 98 per the ISO 17442-1:2020 Annex A check-digit algorithm, which converts any letters in positions 1-18 to digits A=10 ... Z=35 before the mod 97-10 computation; for an all-zero base the conversion is a no-op); implementations MUST NOT use a real third-party LEI in documentation or test data. Where no LEI is allocated and the Deployer is a US entity, an Employer Identification Number (EIN) issued by the United States Internal Revenue Service or a Central Index Key (CIK) issued by the United States Securities and Exchange Commission MAY be used, expressed as the bare numeric string. Decentralized Identifiers () MAY be used otherwise. Implementations MUST treat the value as opaque on verification; identifier resolution is out of scope for this profile. issuer_id values MUST be bare identifiers without a scheme prefix where the scheme is unambiguous from the value's syntactic form. An LEI is the 20-character alphanumeric string defined by and is self-identifying through its length and check-digit structure; implementations MUST emit the bare 20-character LEI without a lei: or other scheme prefix. EINs and CIKs are likewise emitted as the bare numeric string. Decentralized Identifiers () carry their own scheme prefix (did:) as defined by the DID specification and that prefix is intrinsic to the identifier syntax rather than an added scheme tag. The same kid-equals-issuer_id invariant requires signature.kid to be the bare identifier in the same form. The worked example in uses the bare 20-character placeholder LEI 00000000000000000098; conformant cloud emitters and SDK clients MUST match this form on the wire.

payload_digest (OPTIONAL upstream, REQUIRED in this profile) REQUIRED for Compliance Receipts. The value MUST follow the upstream object form (hash, size, optional preview) defined in Section 2.2 of ; this profile does not redefine the wire shape. The associated payload that this digest covers MUST be retained for the period mandated by the most restrictive applicable regime in Sections 5 and 6 of this document. Implementations MUST NOT discard the underlying payload while a receipt that references it is still within its retention window.

action_ref (OPTIONAL upstream, REQUIRED in this profile) REQUIRED for Compliance Receipts. The value is a SHA-256 hash of the canonical Action representation as defined in . This profile uses action_ref as the primary join key for cross-engine reconstruction during an audit.

sandbox_state (OPTIONAL upstream, REQUIRED for High-Risk in this profile) REQUIRED for receipts produced by High-Risk AI Systems under either or . Upstream defines sandbox_state as an OS-level containment status and restricts the value to one of enabled, disabled, or unavailable; this profile inherits that enumeration unchanged. A Deployer that operates a High-Risk AI System and produces a stream of receipts in which sandbox_state is consistently disabled SHOULD treat that stream as a finding under the applicable risk-management documentation requirement (Article 9 of for the Provider's risk management system, with which a Deployer operating per Article 26(1) is required to be consistent; Section 6-1-1703(2) of the Colorado Revised Statutes) and document the rationale in the Audit Pack metadata.

iteration_id (OPTIONAL upstream, REQUIRED for multi-step in this profile) REQUIRED for multi-step agent workflows. The value MUST be stable across all receipts emitted within the same logical task or session so that a regulator can reconstruct the full chain of Actions. iteration_id is distinct from the upstream session_id field defined in Section 3.1.1, which is an opaque MCP session identifier. A Compliance Receipt MAY carry both: session_id for MCP-session correlation and iteration_id for logical-task correlation.

Decision Receipt Fields (type protectmcp:decision) The decision field value MUST be allow, deny, rate_limit, or observation. Implementations using a different internal vocabulary (e.g. permit for allow) MUST normalise on emission and on Audit Pack export. The observation value records that an Action was observed and the receipt was signed without any policy evaluation having taken place; it is the regulator-honest alternative to emitting allow when no policy matched, and MUST NOT appear in a receipt of type protectmcp:decision. A producing system that has not evaluated a policy for an Action MUST either refuse to issue a Compliance Receipt for that Action or MUST emit the receipt with type protectmcp:lifecycle and decision observation; in the latter case the upstream policy_decision internal field, if present in the producing system's internal vocabulary, takes the literal value none, which the emitter MUST map to observation on the wire. Verifiers MUST reject a Compliance Receipt that carries decision observation together with type protectmcp:decision; conversely, a Compliance Receipt of type protectmcp:lifecycle MAY carry decision observation in addition to the other three vocabulary values. The policy_digest requirement of applies to observation receipts in the form of a digest of the producing system's "no policy matched" sentinel policy artefact, which the Deployer MUST retain alongside its other policy artefacts for the applicable retention window. The upstream tool_name field (REQUIRED in Section 3.1.1) is REQUIRED for Compliance Receipts of type protectmcp:decision.

reason (OPTIONAL upstream, REQUIRED for deny/rate_limit in this profile) REQUIRED for Compliance Receipts where decision is deny or rate_limit. The value MUST be a machine-readable reason code drawn from a vocabulary documented in the Deployer's Audit Pack metadata.

policy_digest (OPTIONAL upstream, REQUIRED in this profile) REQUIRED for Compliance Receipts. The value MUST be of the form sha256:<hex> and MUST reference a policy artefact that the Deployer retains for the applicable retention window. Verifiers MUST reject Compliance Receipts whose policy_digest does not resolve in the Audit Pack.

scanner_decisions (OPTIONAL) An OPTIONAL scanner_decisions field MAY appear in a Compliance Receipt of type protectmcp:decision. When present, its value MUST be a JSON array of objects, where each object records the outcome of one content scanner plug-in that ran during policy evaluation. Each object has the following members:

• scanner_id (string, REQUIRED): stable identifier for the scanner plug-in (e.g. presidio, llm-guard, cedar).
• scanner_version (string, REQUIRED): vendor-reported version string for the scanner instance that produced the outcome.
• scanner_decision (string, REQUIRED): one of allow, scan_blocked, or observation; matches the wire vocabulary of the parent decision field where applicable.
• latency_ms (integer, OPTIONAL): wall-clock time in milliseconds the scanner took to produce the outcome.
• signature (string, OPTIONAL): vendor-supplied detached signature over the scanner outcome, used by Deployers that require non-repudiation per scanner instance.

The field is OPTIONAL because Deployers running a single in-process scanner (the common starter deployment) gain no auditing value from echoing the outcome and would only pay an envelope-size cost. The per-scanner signature member is OPTIONAL for the same reason: cryptographic non-repudiation across multiple independent scanner vendors is an Enterprise-tier concern and not mandated at this profile revision. Verifiers MUST tolerate the absence of scanner_decisions and MUST NOT infer that no scanners ran from its absence.

Hash-Chain Linkage (OPTIONAL upstream, REQUIRED in this profile) Upstream Commitment Mode introduces previousReceiptHash as part of an optional extension. This profile makes the linkage REQUIRED. Implementations MUST emit a previousReceiptHash field, populated per the digest-scope rule of Section 5.7 of : the lowercase hex encoding of SHA-256 over the canonical signing-input bytes of the immediately prior receipt emitted by the same issuer_id, where the canonical signing-input bytes are the JCS-canonical serialization () of the predecessor's signed payload object (the same bytes the predecessor's cryptographic signature covers), NOT the envelope object that additionally includes the signature or anchors top-level keys. The first receipt in a chain MUST set this field to the all-zero SHA-256 value (this profile's stipulation; Section 5.7 specifies only the digest scope of subsequent links). JSON key is the literal previousReceiptHash (camelCase, case-sensitive); snake_case aliases MUST NOT appear on the wire. Rationale for the payload-bytes (signing-input) digest scope rather than the envelope-including-signature digest scope: the chain layer's purpose is to make after-the-fact alteration of the predecessor's signed content detectable, and the predecessor's signed content is exactly the bytes its signature covers (the canonical payload object). Digesting those bytes binds the chain to what A actually attested to, is recomputable offline from the predecessor's payload alone, and matches Section 5.7 of . The chain does not need to bind the predecessor's signature value directly because the predecessor's signature is verified independently under , and cross-agent envelope integrity (where binding the peer's signature value matters) is the role of counterparty_binding per , which digests at the envelope-including-signature scope precisely because the peer signature is the load-bearing artefact in the cross-agent case. Implementations that previously digested the envelope-including-signature object MUST migrate to the signing-input scope before emitting chained receipts under this profile; verifiers MUST recompute under the signing-input scope when checking previousReceiptHash. Each issuer MUST maintain a single linear per-agent chain. When one agent identity emits receipts from multiple concurrent execution paths (for example parallel tool calls dispatched within a single agent loop, or fan-out work performed by a thread pool inside one issuer), the issuer MUST serialize emission through a single predecessor pointer at a time: each newly emitted receipt's previousReceiptHash MUST resolve to the SHA-256(JCS(receipt)) of the immediately prior receipt emitted by that same issuer_id, taken in emission order, regardless of which concurrent execution path produced it. Parallel sub-chains within one agent identity (for example, a per-receipt chain_id discriminator that would partition one issuer's stream into multiple independently advancing chains) are NOT defined by this profile. An issuer that requires parallel sub-chains MUST express each parallel path as a distinct agent identity, with its own issuer_id value, its own signing key, and its own per-agent chain rooted at the all-zero SHA-256 genesis value. Rationale: deterministic verification of the chain segment covering an audit window, as required by the regime bindings of Sections 5 and 6 (in particular , , and ), depends on a single linear total order over the receipts emitted under each agent identity; a verifier reconstructing the chain from a regulator-supplied issuer_id needs that ordering to be well-defined without out-of-band metadata.

Anchoring (No Upstream Equivalent) lists Sigstore Rekor in its Implementation Status appendix as an OPTIONAL temporal anchor. This profile imposes a normative anchoring requirement. Compliance Receipts MUST be anchored. An anchor is an timestamp token covering the signed envelope, an commitment covering the envelope, or both; implementations SHOULD emit both forms. For both anchor types, the bytes committed are SHA-256(JCS(envelope_minus_anchors)), where envelope_minus_anchors is the wire envelope object with the anchors top-level key removed prior to canonicalization, leaving the two-key object {payload, signature}. The anchors key MUST be removed from the object, not set to null or to an empty array; these produce different JCS output and break interoperability (mirroring the upstream Section 5.6 stripping rule). The anchor thereby binds payload and signature without being self-referential. The anchor evidence MUST be retained alongside the receipt for the applicable retention window. Verifiers MUST reject Compliance Receipts that lack at least one valid anchor. An anchor MAY be attached after issuance if the receipt is persisted with an unambiguous pending marker and the anchor lands within a documented bound. For , this profile imposes a 7-day deadline; this is a profile-imposed bound, not a property of the OpenTimestamps protocol, whose calendar-to-block upgrade time depends on the calendar operator's publication interval. tokens MUST be obtained synchronously. A verifier MUST treat a pending receipt as non-conformant once the bound elapses. The anchor MAY cover an aggregate of receipts (for example, a Merkle root over a batch) rather than each receipt individually, provided that the inclusion proof linking the receipt to the aggregate is retained alongside the receipt and the aggregate anchor. Where the anchor type is , the full TimeStampResp DER bytes MUST be retained, sufficient for offline verification by a holder with access to the TSA's published public key. Time-stamp tokens carrying ESSCertIDv2 per MUST be accepted by Compliance Verifiers. Where the anchor type is , the upgrade from the initial calendar attestation to the Bitcoin block attestation MUST be completed within the 7-day profile-imposed bound, and the upgraded proof MUST be retained for the applicable retention window per the second paragraph of this section. Each entry in the top-level anchors array is an object with the following members.

type:

REQUIRED string discriminator. MUST be one of rfc3161 or opentimestamps.

value:

REQUIRED on every anchor entry. The anchor token bytes, base64-encoded. For rfc3161 the value is the base64 encoding of the full TimeStampResp DER bytes (sufficient for offline cryptographic re-verification by a holder with access to the TSA's published public key). For opentimestamps the value is the base64 encoding of the OpenTimestamps proof blob (the .ots serialization). A verifier MUST cryptographically re-verify the anchor against the signed envelope using these bytes per ; anchor entries served without value MUST NOT be reported as anchor_valid_*=true.

status:

OPTIONAL informational string. When present, MUST be one of anchored (the anchor has reached its final attestation state: an token has been obtained, or an commitment has upgraded to its Bitcoin block attestation), pending (the anchor has been requested but the final attestation state has not yet been reached, e.g. an OpenTimestamps commitment that has been submitted to a calendar but has not yet upgraded to a Bitcoin block within the 7-day bound of this section), or failed (the anchor submission was attempted and did not produce a usable attestation, e.g. a TSA returned an error response or an OpenTimestamps calendar refused the commitment). status is operational metadata; a verifier MUST NOT derive cryptographic validity from status alone, and MUST always re-verify the value bytes per .

bitcoin_block:

OPTIONAL informational string. The Bitcoin block hash at which an commitment was anchored. Present only on entries with type=opentimestamps and status=anchored; absent on rfc3161 entries and on pending or failed OpenTimestamps entries. bitcoin_block is operational metadata; a verifier MUST NOT derive cryptographic validity from bitcoin_block alone, and MUST always re-verify the value bytes against the OpenTimestamps proof per .

Extension Fields This profile defines three extension fields that MAY appear in the signed payload alongside the fields defined by . The first two, risk_class and incident_class, are defined in this section. The third, counterparty_binding, is defined in .

risk_class:

A vocabulary term identifying the risk classification of the Action under the Deployer's risk management documentation. The vocabulary MUST be referenced in the Audit Pack metadata. Where the Deployer operates under , the documentation is the Provider's Article 9 risk management system as referenced via the instructions for use under Article 26(1); where the Deployer operates under , the documentation is the Section 6-1-1703(2) risk management policy and program.

incident_class:

A vocabulary term identifying the incident classification of the Action under the applicable regime: an ICT-related incident under , with classification criteria in and the canonical reporting enumeration of Annex II data glossary, field 3.23 (Type of the incident) of (verifiers MUST resolve the canonical values from the regulation directly); a Cybersecurity Event under 23 NYCRR 500.1(f) (or, where the Section 500.17(a) reporting threshold is met, a Cybersecurity Incident under 23 NYCRR 500.1(g)) for Covered Entities of ; a Covered Cyber Incident under once the final rule takes effect; or a security incident under 45 CFR 164.304 for Covered Entities of . Implementations MAY refine the set, provided the flattened mapping in the Audit Pack manifest () projects each refinement to the applicable canonical category for each in-scope regime.

risk_class MUST be encoded as a JSON string. incident_class MUST be encoded as a JSON string drawn from the canonical vocabulary referenced in the Audit Pack, OR as a JSON array of such strings to preserve cross-regime classification (for example, a single Action that is both a DORA ICT-related incident and a CIRCIA Covered Cyber Incident, or both a NYDFS Cybersecurity Incident and a CIRCIA Covered Cyber Incident). Both extension fields appear inside the signed payload object and are therefore covered by the upstream Section 5.6 signature scope. Both fields are OPTIONAL at the syntactic level but MAY be REQUIRED by the regime bindings in Sections 5 and 6 of this document. Implementations MAY define additional extension fields. Such fields MUST NOT collide with names defined by or by this document. Implementations defining extension fields SHOULD register them in the registry described in .

Counterparty Binding This section is normative. counterparty_binding is an in-payload object an acknowledging agent ("B") emits to carry a cryptographic digest of the full signed envelope of an originating agent ("A"). It provides cross-agent byte-equality evidence when a shared intermediary sits between two honest agents and the per-agent hash chains of validate independently regardless of whether B's observed bytes equal A's signed bytes. action_ref is a correlation anchor, not a cryptographic binding ( Section 2.2); counterparty_binding moves the evidence onto B's own COSE or JWS signature, which the verifier already trusts.

Wire Shape The field MUST appear inside the signed payload object. It MUST NOT appear in unprotected COSE or JWS header parameters, or in external_aad per Section 4.3 when the receipt is used for audit (external_aad is permissible only in transport-optimized modes out of scope for Compliance Receipts). For COSE-framed receipts the field sits inside the COSE_Sign1 or COSE_Sign payload per Section 4.1; for JWS-framed receipts it is a top-level claim per . The field is an object with the following members.

envelope_hash:

REQUIRED string. Base64-encoded SHA-256 digest computed over A's entire serialized signed envelope, including A's signature bytes. The digest input is framing-specific. JSON-framed (this profile's default for receipts that are not transported under COSE or JWS, and mandatory-to-implement for any conformant Compliance Receipt implementation): the JCS-canonical UTF-8 byte sequence of A's signed envelope JSON object per , where the envelope is the three-key object {"payload": <signed payload object>, "signature": <signature object with alg, kid, sig members>, "anchors": <array of anchor objects, OPTIONAL>}. The payload object carries the signed fields A emitted (including type, issuer_id, issued_at, action_ref, payload_digest, previousReceiptHash, decision, and any extension fields under ); the signature object carries the algorithm identifier, key identifier, and base64- or base64url-encoded signature bytes exactly as A emitted them. B MUST NOT re-canonicalize A's payload or strip the anchors array before computing the digest. COSE-framed: the full COSE_Sign1 or COSE_Sign byte string per Section 4.2 (deterministic encoding). JWS-framed: the full JWS Compact Serialization (header.payload.signature) after payload canonicalization per . The digest algorithm is SHA-256 (mandatory-to-implement); the encoding MUST be standard base64 per Section 4 on emission, OR base64url per Section 5 where the surrounding transport requires URL-safe encoding; verifiers MUST accept both alphabets and MUST normalise to a single alphabet (typically standard base64) before byte-comparing to a recomputed value. Including A's signature in the digest scope binds the signed-over content of A's receipt at the envelope level and prevents an intermediary that re-signs A's claims with a different key from escaping detection. The framing in which A's envelope was emitted MUST be preserved through B's binding: the value of envelope_hash is framing-specific because JCS-canonical JSON (UTF-16 code-unit lexicographic key ordering per ), COSE deterministic encoding (length-then-byte map-key ordering per Section 4.2), and JWS Compact Serialization with JCS-canonical payload (UTF-16 code-unit lexicographic ordering per ) produce different byte sequences from the same semantic payload-and-signature, and the three framings therefore yield different envelope_hash values for the same underlying receipt. A transcoding intermediary that re-frames A's envelope (JSON to COSE, COSE to JWS, or any other pairing) changes the digest input and MUST be treated as a tampering event by the verifier; verifiers MUST NOT reframe an envelope before recomputing envelope_hash. Future revisions MAY extend to additional digest algorithms drawn from the digest algorithm registry; implementations that require algorithm negotiation SHOULD carry the algorithm identifier out of band in the Audit Pack manifest rather than in the wire field.

receipt_ref:

REQUIRED opaque content-addressed locator the verifier resolves through the Audit Pack or a Deployer-published index to A's full signed envelope. The value is an opaque string from the verifier's perspective; producers MAY use any stable identifier scheme (URI, content-addressed digest, opaque database id) so long as the Audit Pack resolution layer returns the correct envelope bytes. Future profiles (for example, a SCITT-style inclusion-proof profile under Section 4.2 extension semantics) MAY layer on this field.

expect_ack_from:

OPTIONAL string. The expected acknowledging-party identifier, expressed as a kid or issuer_id value matching the same bare-identifier form required by . When present, the field declares which acknowledging party A or the producer expects to sign over this receipt's bytes; a verifier MAY use expect_ack_from to cross-check that the acknowledging receipt's kid matches the expected identifier. Verifiers MUST NOT reject solely on absence of an acknowledging receipt; absence is a liveness-loss signal observable through Audit Pack metadata rather than a non-conformance condition on the current receipt.

transport_label:

OPTIONAL string (mcp, bus, orchestrator, http). Operational only; verifiers MUST NOT derive trust from this label.

The COSE form follows the same member set under deterministic CBOR map ordering per Section 4.2.

Emitter Behaviour B SHOULD emit counterparty_binding when any of the following hold: A's signing request flagged the action as requiring acknowledgment (for example, by populating an expect_ack_from list); the Deployer's risk management documentation requires bilateral byte-binding; or B is operating under the guidance of . B MUST compute envelope_hash over the exact byte stream it received and accepted, not over a re-canonicalization at B; re-canonicalizing at B masks intermediary tampering whenever the tampered bytes canonicalize to the same payload object, which is the threat case this section addresses. Where one acknowledgment receipt confirms envelopes from N originators, the field MAY be an array of objects; pairwise bindings cannot prove all N originators emitted identical bytes (see ).

Verifier Behaviour A Compliance Verifier processing a receipt carrying counterparty_binding MUST, in addition to , resolve receipt_ref through the Audit Pack or a Deployer-published index to A's full signed envelope, recompute the SHA-256 digest of that envelope under the scope rule of , base64-encode the result, and compare to envelope_hash. A non-resolving receipt_ref or a digest mismatch MUST cause the acknowledging receipt to be reported non-conformant; liveness loss at A MUST NOT be silently treated as success. Where expect_ack_from is present, the verifier SHOULD additionally check that the acknowledging receipt's signature.kid matches the declared identifier (under the bare-identifier form required by ); mismatch SHOULD be reported as an axis flag rather than as outright non-conformance because the field is OPTIONAL. The Deployer or Audit Pack producer MUST retain A's signed envelope for at least as long as any acknowledging receipt binding it remains within retention under Sections 5 and 6. For chains of three or more agents, this profile defaults to pairwise bindings; multi-signer co-presence under Section 4.1 is OPTIONAL, and verifiers MUST NOT treat a co-signed envelope as a substitute for a pairwise binding chain.

European Union Bindings

EU AI Act Article 12 Binding Each subsection cites the operative phrase of Article 12 and binds it to the receipt field that satisfies it.

Article 12(1), automatic recording of events Article 12(1) requires High-Risk AI Systems to technically allow for the automatic recording of events (logs) over the lifetime of the system. The signed-receipt format provides one mechanism that satisfies that logging capability; alternative mechanisms remain valid. Where this profile is chosen, a Compliance Receipt SHOULD be produced for every Action against an external resource, and a configuration change that disables receipt generation SHOULD be recorded as a protectmcp:lifecycle Compliance Receipt. Implementations MAY emit at finer or coarser granularity so long as the log set, taken together, satisfies Article 12(2)(a) through (c).

Article 12(2)(a), identifying situations that may result in the high-risk AI system presenting a risk within the meaning of Article 79(1) or in a substantial modification The combination of type, decision, reason, and policy_digest MUST be sufficient for an auditor to identify, by query alone, receipts that correspond to risk situations enumerated in the Deployer's risk management documentation. Where the Deployer classifies an Action as risk-bearing, the receipt MUST carry a risk_class extension field.

Article 12(2)(b), facilitating the post-market monitoring referred to in Article 72 The hash-chain linkage required by satisfies post-market monitoring traceability. The chain head MUST be made available to the Provider and to the competent authority on request.

Article 12(2)(c), monitoring the operation of high-risk AI systems referred to in Article 26(5) Any change to the policy artefact referenced by policy_digest MUST produce a new digest. A change in policy_digest between two otherwise-comparable Actions may be examined by the Deployer or by a regulator as a candidate substantial-modification event under Article 43, and MUST be retained at least as long as the longest receipt in the chain that references either digest.

Retention Article 12 itself sets no retention period; the operative deployer floor is Article 26(6) ("at least six months"). The parallel provider floor in Article 19(1) sets the same six-month minimum on Providers; this profile's retention bindings are written from the Deployer perspective, and a Provider that wishes to use Compliance Receipts as its Article 19(1) record SHOULD adopt the Deployer floor explicitly through a separate Provider-role binding (deferred to a future revision). EU AI Act Article 26(6) requires six months of logs (interpreted as 184 days when expressed as a day-count floor for hash-chain anchoring intervals). The normative requirement on Compliance Receipts is: implementations MUST retain receipts until the later of (a) the day six calendar months after the date of the Action, computed calendar-arithmetically per duration arithmetic, and (b) any longer Union or national law floor. The six-month period is read calendar-month-wise (the receipt expiry day is the same day-of-month six months later, with end-of-month rollover where the target month is shorter), not as a fixed day count. The 184-day figure (the maximum number of days in any rolling six-calendar-month window, worst case Aug-Jan, 31+30+31+30+31+31) is informative only; it expresses a safe day-count floor for hash-chain anchoring intervals and Audit Pack export windows where calendar-arithmetic is impractical at the producer layer. A Deployer that retains receipts strictly under the calendar-month rule above satisfies Article 26(6); a Deployer that uses 184 days as an internal day-count overestimate also satisfies it. The 183-day pick mentioned in earlier revisions is not endorsed by this profile: under a rolling six-calendar-month window 1 August to 31 January spans 184 days and 183 days is one day short. Where the Deployer is also a Financial Entity, the sectoral floor in applies.

EU AI Act Article 26 Binding

Article 26(1), in accordance with the instructions for use policy_digest MUST resolve through to a retained artefact (machine check). The Deployer SHOULD demonstrate consistency with the Provider's instructions for use (process check). Inability to perform the machine check is presumed non-compliance.

Article 26(2), assign human oversight For any Action whose decision is allow and which the Deployer's risk management documentation marks as requiring human oversight, the Deployer MUST ensure that the receipt is either reviewed by a designated natural person within the period required by national law, or that a follow-on protectmcp:lifecycle Compliance Receipt records the absence of such review with a reason code. Both records MUST themselves be Compliance Receipts. This profile addresses the trigger and record of oversight; the competence, training, authority, and necessary support of the reviewer required by Article 26(2) remain the Deployer's separate responsibility.

Article 26(5), monitor the operation A Deployer MUST be able to produce an Audit Pack covering any contiguous time window since the High-Risk AI System became operational.

Article 26(6), keep the logs for at least six months Compliance Receipts under this binding MUST be retained for at least the period stated in . Where the Deployer is also a Financial Entity, the longer sectoral floor in applies.

DORA Article 17 Binding

Article 17(1), ICT-related incident management process A Compliance Receipt produced inside a Financial Entity's ICT environment may serve as the canonical record of an Action that triggered an ICT-related incident. action_ref MUST be carried into the Financial Entity's incident workflow as the primary correlation key.

Article 17(2), record all ICT-related incidents and significant cyber threats The hash chain required by supports the recording obligation of Article 17(2) by making after-the-fact alteration of recorded incidents detectable. The Financial Entity MUST be able to produce, on request, the chain segment covering the period of an incident, together with the anchor evidence that fixes the chain to wall-clock time.

Article 17(3)(b), establish procedures to identify, track, log, categorise and classify ICT-related incidents For Actions identified as part of an ICT-related incident, the producing system MUST emit incident_class. The classification criteria are those set out in Article 18(1) of , with further specification in . The canonical reporting enumeration to which incident_class flattens is bound by Annex II field 3.23 of (see ). Implementations MUST publish a flattened mapping in the Audit Pack manifest as required by .

Retention Article 17 of does not itself set a uniform numeric retention floor. The five-year (1827-day) figure used by this profile derives from sectoral instruments that overlap DORA-scoped Financial Entities. Investment firms keep records of all services, activities and transactions under Article 16(6) of , with Article 72 and Annex I of fixing the form and content of those records. The explicit five-year retention period in is set by Article 16(7) for records of telephone conversations and electronic communications, kept for a period of five years and, where requested by the competent authority, for a period of up to seven years. Records of customer due diligence and of transactions under Article 40 of are kept for five years after the end of the business relationship. The AMLD record-keeping regime is superseded, in respect of record retention, by Article 77 of from 10 July 2027, which preserves the five-year floor and adds a case-by-case extension up to a further five years where the competent authority so requires. Implementations operating across the AMLD-to-AMLR transition MUST satisfy whichever instrument is in force on the date of the Action. Compliance Receipts MUST be retained for the period required by applicable Union or national law; where a sectoral floor applies, retention MUST equal or exceed the longest applicable floor. Absent a more specific rule, this profile RECOMMENDS 1827 days from the date of the Action (the worst-case rolling five-calendar-year window contains two leap days, so 1827 days satisfies "five years" regardless of the calendar years over which the window falls). Anchor evidence MUST be retained for the same period. Verification keys whose lifetime expires within the retention window MUST have their public components retained so that historical signatures remain verifiable.

United States Bindings

NIST AI RMF Binding is a voluntary framework. Adoption of this profile, on its own, does not establish conformity with the AI RMF; it provides a tamper-evident receipt substrate that an AI RMF program can use as evidence under the MEASURE function and as a structured input to the GOVERN, MAP, and MANAGE functions. applies the AI RMF functions to generative AI; the profile bindings below apply to generative and non-generative AI agent deployments alike unless explicitly noted.

GOVERN function The GOVERN function requires that organizations document AI policies and procedures. The combination of policy_digest and the Audit Pack manifest provides a machine-readable binding between every Action and the policy artefact in force at the time of the Action. A change to the policy artefact MUST produce a new policy_digest value (per ); the Audit Pack therefore records every policy change in a tamper-evident manner.

MAP function The MAP function requires that the context, capabilities, and risks of an AI system be characterised. The combination of type, tool_name, action_ref, and iteration_id SHOULD be sufficient for an auditor to reconstruct the operational context of any Action without dereferencing the underlying payload.

MEASURE function The MEASURE function requires that AI risks and impacts be analysed and tracked over time. The hash-chain linkage required by provides tamper-evident continuity of the receipt stream over the AI system's operational lifetime, satisfying the traceability prerequisite of MEASURE.

MANAGE function The MANAGE function requires that AI risks be prioritised and acted upon based on projected impact. The risk_class extension field carries the Deployer's risk classification of the Action; together with decision, reason, and policy_digest, it supports prioritisation and incident response without requiring the verifier to re-derive risk from the underlying payload.

Colorado AI Act (SB 24-205) Binding imposes deployer obligations effective June 30, 2026 (per Senate Bill 25B-004, which postponed the original February 1, 2026 effective date). The Act regulates the deployment of High-Risk AI Systems and the prevention of algorithmic discrimination.

Section 6-1-1703(2), risk management policy and program Section 6-1-1703(2) requires deployers to implement a risk management policy and program for the High-Risk AI System. policy_digest MUST resolve through to the deployer's risk management policy artefact in force at the time of the Action. Where the Deployer classifies an Action as risk-bearing under that policy, the receipt MUST carry a risk_class extension field.

Section 6-1-1703(3), impact assessment Section 6-1-1703(3) requires deployers to complete an impact assessment annually and within 90 days after any intentional and substantial modification of the High-Risk AI System. The combination of type, policy_digest, and previousReceiptHash MUST be sufficient for an auditor to identify, by query alone, the receipts that span the period covered by an impact assessment, including any policy changes within that period.

Section 6-1-1703(7), notice of algorithmic discrimination Where a Deployer determines that a High-Risk AI System has caused or is reasonably likely to have caused algorithmic discrimination, the producing system SHOULD record that determination as a protectmcp:lifecycle Compliance Receipt naming the determination, the affected receipts by action_ref, and the policy or risk-management response with a reason code.

Texas Responsible AI Governance Act (HB 149) Binding takes effect January 1, 2026. The Act adopts an intent-based liability framework for the development and deployment of AI systems and provides a safe harbor at Section 552.105(e)(2)(D) of the Texas Business and Commerce Code for organisations that substantially comply with the most recent version of , or another nationally or internationally recognized risk management framework for AI systems, and operate an internal review process.

Safe-harbor evidentiary support Where a Deployer relies on the safe-harbor provision of HB 149 by substantially complying with , the Audit Pack MAY be presented as evidence of that compliance. The bindings of apply, with the additional Generative AI Profile bindings of .

Prohibited-use detection Receipts whose decision is deny with a reason code drawn from a vocabulary documenting the Act's prohibited-use categories under Section 552.052 of the Texas Business and Commerce Code added by HB 149 (incitement or encouragement of physical self-harm including suicide, harm to another person, or engagement in criminal activity) MUST be retained for the period stated in or the longer period required by Texas law, whichever is greater.

HIPAA Security Rule Binding (45 CFR Part 164, Subpart C) applies to Covered Entities (HIPAA) that handle electronic protected health information. The bindings below apply only to receipts whose underlying Actions reference electronic protected health information.

45 CFR 164.312(b), audit controls 45 CFR 164.312(b) requires implementation of "hardware, software, and/or procedural mechanisms that record and examine activity in information systems that contain or use electronic protected health information". The combination of type, action_ref, tool_name, and the hash-chain linkage required by satisfies the recording requirement; the verification rules of satisfy the examination requirement.

45 CFR 164.316(b)(2), six-year retention 45 CFR 164.316(b)(2) (and in particular the subparagraph 164.316(b)(2)(i)) requires that the documentation required by 45 CFR 164.316(b)(1) be retained "for 6 years from the date of its creation or the date when it last was in effect, whichever is later". The audit-log content produced under 45 CFR 164.312(b) is not itself documentation required by 164.316(b)(1); the Security Rule does not set an explicit retention floor for individual audit-log records. By analogy with the six-year floor that 164.316(b)(2) places on the policies and procedures that govern audit-log generation, this profile applies the same six-year floor to Compliance Receipts whose underlying Actions reference electronic protected health information. Records covered by the HIPAA Security Rule audit-trail retention MUST be retained for six years from the date of creation or the date when last in effect, whichever is later, per 45 CFR 164.316(b)(2). This profile expresses that floor as 2192 days from the later of (a) the date of the Action and (b) the date the policy artefact referenced by policy_digest ceased to be in effect: 2192 is the maximum number of days in any rolling six-calendar-year window (worst case spans two leap days, e.g. 2024-2030 contains February 29 of 2024 and 2028, yielding 6*365+2 = 2192 days). The six-year analogy floor is grounded in 164.316(b)(2); a Covered Entity that retains receipts strictly under 164.316(b)(2)(i) bound only to the policy artefact's creation-or-cessation date MAY do so when no longer audit-log floor is established by separate Union, state, or sectoral law. Verification keys whose lifetime expires within the retention window MUST have their public components retained so that historical signatures remain verifiable.

NYDFS Cybersecurity Regulation Binding (23 NYCRR Part 500) applies to Covered Entities (NYDFS) operating under New York Banking, Insurance, or Financial Services Law. The bindings below apply only to receipts produced by such Covered Entities.

23 NYCRR 500.6, audit trail 23 NYCRR 500.6(a) requires Covered Entities to securely maintain systems that, to the extent applicable and based on its risk assessment, (1) are designed to reconstruct material financial transactions, and (2) include audit trails designed to detect and respond to cybersecurity events that have a reasonable likelihood of materially harming any material part of the normal operations of the Covered Entity. The hash chain required by together with the anchor evidence required by satisfies the tamper-evidence prerequisite of the audit-trail obligation.

23 NYCRR 500.17, notices to superintendent 23 NYCRR 500.17(a)(1) requires that "Each covered entity shall notify the superintendent electronically in the form set forth on the department's website as promptly as possible but in no event later than 72 hours after determining that a cybersecurity incident has occurred at the covered entity, its affiliates, or a third-party service provider." The reporting trigger is a Cybersecurity Incident under 23 NYCRR 500.1(g), not any Cybersecurity Event under 500.1(f). For Actions identified as part of such an Incident, the producing system MUST emit incident_class with a value indicating Cybersecurity Incident under 23 NYCRR 500.1(g), and the Covered Entity MUST be able to produce, on request, the chain segment covering the period of the Incident together with the anchor evidence that fixes the chain to wall-clock time.

23 NYCRR 500.6 retention 23 NYCRR 500.6(b) requires that "Each Covered Entity shall maintain records required by this section for not fewer than five years." The five-year floor applies uniformly to records required by paragraph (a)(1) (designed to reconstruct material financial transactions) and to records required by paragraph (a)(2) (audit trails designed to detect and respond to cybersecurity events that have a reasonable likelihood of materially harming any material part of the normal operations of the Covered Entity). Compliance Receipts produced under this binding MUST be retained for at least 1827 days from the date of the Action (the worst-case rolling five-calendar-year window contains two leap days).

SEC Broker-Dealer Recordkeeping Binding (17 CFR 240.17a-4) applies to Broker-Dealers, Security-Based Swap Dealers, and Major Security-Based Swap Participants. The bindings below apply only to receipts produced inside such entities.

17 CFR 240.17a-4(f), electronic recordkeeping system The November 3, 2022 amendments to 17 CFR 240.17a-4 (compliance date May 3, 2023) added an audit-trail alternative to the prior write-once-read-many (WORM) electronic recordkeeping requirement. The audit-trail alternative requires that the electronic recordkeeping system permit the recreation of an original record if it is modified or deleted. The hash-chain linkage required by together with the retention rule in and the anchor evidence required by satisfies the audit-trail alternative when the Compliance Receipt is the system-of-record for a regulated record.

17 CFR 240.17a-4(a) and (b) retention 17 CFR 240.17a-4(a) requires preservation of certain records for not less than 6 years, the first two years in an easily accessible place. 17 CFR 240.17a-4(b) requires preservation of a different list of records for not less than three years, the first two years in an easily accessible place. Compliance Receipts that constitute or support a record listed in 17 CFR 240.17a-4(a) MUST be retained for at least 2192 days from the date of the Action, applying the same six-year worst-case methodology as ; receipts that constitute or support a record listed only in 17 CFR 240.17a-4(b) MUST be retained for at least 1096 days from the date of the Action (the worst-case rolling three-calendar-year window contains one leap day). Where both apply, the longer period applies.

CIRCIA Binding (Cyber Incident Reporting for Critical Infrastructure Act of 2022) requires Covered Entities (CIRCIA) to report Covered Cyber Incidents to the Cybersecurity and Infrastructure Security Agency within 72 hours of reasonable belief that the incident has occurred, and to report ransom payments within 24 hours. The reporting obligations take effect upon publication of the final rule. Pending publication, the bindings below apply on a voluntary basis.

Covered Cyber Incident reporting support For Actions identified as part of a Covered Cyber Incident, the producing system MUST emit incident_class with a value indicating Covered Cyber Incident under . The Covered Entity MUST be able to produce, on request, the chain segment covering the period of the incident together with the anchor evidence that fixes the chain to wall-clock time.

Records related to a Covered Cyber Incident report Section 2242(a)(4) of the Homeland Security Act of 2002, as enacted by and codified at 6 U.S.C. 681b(a)(4), requires Covered Entities to preserve data relevant to a Covered Cyber Incident or ransom payment in accordance with procedures established in the final rule. CISA's notice of proposed rulemaking at 89 FR 23644 (April 4, 2024), proposed Section 226.13(c), proposes a preservation period of not less than two years measured from the submission of the most recently required CIRCIA report (or the date that submission would have been required absent a preservation exception under proposed Section 226.4(a)); this profile uses that proposed floor pending publication of the final rule. Records covered by CIRCIA preservation MUST be retained for two years from the submission of the most recently required CIRCIA report under 6 U.S.C. 681b(c)(2). Compliance Receipts that are referenced in a CIRCIA report or that the Covered Entity reasonably anticipates will be so referenced MUST be retained for the longer of (a) the period established by the final rule and (b) two years from the submission of the most recently required CIRCIA report (or the date that submission would have been required absent a preservation exception), per proposed Section 226.13(c) of the CIRCIA NPRM at 89 FR 23644 (April 4, 2024). Covered Entities MUST NOT measure the retention floor from the date of the underlying Action; an Action detected and reported months later carries a preservation window that runs forward from the report submission date.

Audit Pack Composition This section is informative. It describes the contents of an Audit Pack as introduced in . An Audit Pack contains the following items.

• The set of Compliance Receipts covered by the requested time window, in the canonical envelope form defined by .
• The chain commitments that link the receipts: for each receipt, the value of previousReceiptHash and the recomputed digest of the predecessor envelope.
• The anchor evidence: tokens, OpenTimestamps proofs, or both. Each anchor item MUST be associated, by hash, with the receipt or aggregate it covers.
• The trust anchor metadata that identifies the Deployer or other regulated entity associated with each issuer_id value.
• The verification key material for every kid value present, in a form that does not require online retrieval.
• Vocabularies referenced by reason, risk_class, incident_class, and extension fields, embedded as JSON arrays with a stable identifier. The Audit Pack MUST expose a digest-resolution facility that, given a policy_digest, returns the retained artefact.
• A regime mapping document that names which receipts the producer asserts as evidence under any of the regimes addressed by Sections 5 and 6 of this document (EU AI Act Article 12, EU AI Act Article 26, DORA Article 17, NIST AI RMF, Colorado AI Act, Texas Responsible AI Governance Act, NYDFS Part 500, HIPAA Security Rule, SEC Rule 17a-4, CIRCIA).
• The chain heads valid at the start and end of the time window, signed by the Deployer or other regulated entity.

An Audit Pack MUST itself be signed per the algorithm registry. The manifest MUST include bundle_digest, bundle_signature, bundle_public_key, and algorithm_registry_version. The following manifest-level fields SHOULD appear on an Audit Pack bundle when the underlying receipt stream exposes the corresponding semantics. Each is informative and does not alter the wire shape of individual receipts.

regime_mapping_disclaimer:

String emitted on bundles whose per-receipt regime predicates derive from mapping logic the original producing system did not sign. The value identifies the producer of the mapping, the document version under which it was computed, and a disclaimer that the regime-satisfaction flags are advisory and remain subject to the verifier's own check against Sections 5 and 6.

stale_pending:

Boolean flag set per bundle entry whose anchor evidence is still pending after the bound of (7 days for OpenTimestamps; synchronous for RFC 3161). When true, the bundled receipt is non-conformant per , and a Compliance Verifier consumes the flag to drive anchor_valid_* false in its per-axis report (see ). A verify endpoint over the same bundle SHOULD surface stale_pending in its response.

Verifier Behaviour A verifier conformant to this profile is referred to as a Compliance Verifier.

Mandatory Checks A Compliance Verifier MUST perform all of the following checks before treating a receipt as a Compliance Receipt.

• Verify the signature using the algorithm declared in signature.alg, in accordance with .
• Resolve the verification key through one of the key-distribution mechanisms described in Section 4.3 of (well-known JWK Set or out-of-band distribution), or through Audit Pack trust-anchor metadata. The verifier MUST NOT trust a verification key embedded in the receipt envelope.
• Verify that all fields marked REQUIRED by are present and well-formed.
• Verify the hash-chain linkage by recomputing SHA-256 over the canonical signing-input bytes of the immediately preceding receipt (the JCS-canonical serialization of the predecessor's signed payload object, per ) and comparing the lowercase hex encoding to previousReceiptHash.
• Verify at least one anchor: an token, an commitment, or both. The anchor MUST cover the signed envelope as it appears in the receipt. The verifier MUST cryptographically re-verify the anchor against the signed envelope; presence of anchor metadata without a successful cryptographic check MUST NOT yield "valid".
• Verify the future-skew bound on issued_at per . Past skew MUST NOT cause non-conformance when the receipt is within retention.
• Verify that policy_digest resolves through . A digest computed over a nonced or otherwise mixed-input form (for example, SHA-256(nonce || JCS(artefact))) MUST NOT be treated as policy_digest; the digest scope is the canonical form of the artefact alone. The verifier MUST recompute SHA-256 over the canonical form of the resolved artefact as documented in the Audit Pack manifest, and compare; for JSON artefacts the canonical form is JCS per .

A receipt that fails any of these checks MUST be reported as non-conformant.

Optional Checks A Compliance Verifier MAY additionally perform any of the following.

• Cross-check the issuer_id against an external registry (LEI, EIN, CIK, NPI, GLEIF, or a Deployer-published list).
• Resolve the policy artefact referenced by policy_digest and compare it to a Provider-supplied or Deployer-supplied reference policy.
• Recompute the chain head and compare it to a Deployer-published value.
• Validate incident_class (each element if encoded as an array) and risk_class extension values against the vocabularies referenced in the Audit Pack.

Reporting A Compliance Verifier SHOULD produce a structured per-receipt report that names the regime bindings the receipt satisfies and the outcome of the per-axis checks the verifier performed. The following fields SHOULD be emitted; the axes are independent and consumers MUST NOT collapse them into a single boolean before display.

regimes_satisfied:

Array of short stable regime identifiers drawn from the regimes listed in (for example eu_ai_act, dora, nist_ai_rmf, colorado_ai, texas_traiga, hipaa_security, nydfs_500, sec_17a4a, sec_17a4b, circia). The set is open-ended; consumers MUST treat unknown identifiers as informational. Where this field is inherited from a producer-side mapping in the bundle, the bundle-level regime_mapping_disclaimer of applies.

anchor_valid_ots:

Boolean. true when an anchor is present, has upgraded to the Bitcoin block attestation within the 7-day bound of , and re-verifies cryptographically against the signed envelope; otherwise false (including the case where the only present anchor is RFC 3161).

anchor_valid_rfc3161:

Boolean. true when an token is present, carries an ESSCertIDv2 per where required, and re-verifies cryptographically against the signed envelope; otherwise false.

policy_digest_resolved:

Boolean. true when policy_digest resolved to a retained artefact whose JCS-canonical SHA-256 matches the receipt value, per ; false on absence, resolution failure, or digest mismatch.

duplicate_emission_candidate:

Boolean. true when at least one other receipt sharing action_ref and issuer_id has been observed in the same Audit Pack or in a verifier-maintained index; otherwise false. The axis is informational; this profile does not require verifiers to maintain a cross-receipt index. An absent index MUST report false rather than omit the axis, so consumers learn "no duplicate" from the value and learn "axis unknown" only from the verifier's documented capability set.

A receipt may carry anchor_valid_ots=true and anchor_valid_rfc3161=false (or vice versa) and still satisfy the mandatory anchor check of , which requires only one valid anchor. Where a receipt carries counterparty_binding per , the verifier SHOULD additionally emit the outcome of the check of ; this profile reserves a stable field identifier for that axis pending implementation experience.

Security Considerations This profile inherits all of the security considerations of . The following considerations are specific to the compliance binding.

Tamper Resistance The hash-chain linkage required by provides tamper-evidence at the chain level. An adversary who removes a receipt from the middle of the chain MUST recompute and re-sign every subsequent envelope. The anchor evidence required by binds segments of the chain to wall-clock time, raising the cost of a re-signing attack. Implementations SHOULD anchor at intervals no longer than 24 hours. Implementations operating under DORA Article 17, 23 NYCRR 500.17, or SHOULD anchor at intervals no longer than one hour, given the four-hour initial-notice deadline (with 72-hour intermediate-report and one-month final-report bounds) per DORA Article 17 and the RTS in , the 72-hour notification clock under 23 NYCRR 500.17, and the 72-hour CIRCIA covered-cyber-incident reporting deadline that will apply once the CIRCIA final rule takes effect. A deployment that uses only the signature, without chain linkage and anchoring, can be rolled back by an insider with control of the signing key for the period between the deletion and the next anchor. The MUST clauses of and close that window.

Chain Availability Under Single-Linear Per-Agent Serialization This section is informative. The single-linear per-agent chain requirement of serializes receipt emission for a given issuer_id through a single predecessor pointer. A denial-of-service against the predecessor pointer (database row lock contention, network partition between the emitter and the predecessor store, slow IO, or an adversary deliberately holding the chain-tail lock) therefore bounds the per-agent emission throughput, because every new receipt MUST resolve the digest of the immediately prior receipt before it can be linked. A partial-write failure between predecessor-pointer commit and signature commit can additionally produce chain-head ambiguity if not handled defensively. Issuers SHOULD use a bounded predecessor-lookup timeout (operator-tuned, typically on the order of seconds rather than tens of seconds) and SHOULD emit a structured audit event with type protectmcp:lifecycle and a stable reason code (RECOMMENDED: chain_emission_blocked) when the timeout fires, rather than silently dropping the receipt or stalling caller threads. Issuers SHOULD additionally document a chain-head recovery procedure for crashed emitters: on restart, the issuer re-reads the predecessor row, verifies that no orphan signature exists for the next sequence position, and resumes emission. Operators that require parallel per-issuer throughput beyond what a single linear chain sustains MUST use distinct issuer_id values per parallel path, with separate signing keys and chains rooted at the all-zero genesis value, per the rule in . The threat profile here is availability, not confidentiality or integrity: a successful chain-availability attack delays or drops emission, but it cannot tamper with already-emitted receipts (those are protected by ) and it cannot forge receipts (those are protected by ). The chain_emission_blocked lifecycle receipt is itself a Compliance Receipt and therefore links into the chain once emission resumes, so the gap is detectable rather than silent.

Key Compromise A Compliance Receipt is only as trustworthy as the key that signed it. On suspected compromise of an issuer key, the Deployer MUST publish a revocation notice that names the key, the time of suspected compromise, and the chain head at that time. Receipts signed by the compromised key after the named time MUST NOT be treated as Compliance Receipts. Verifiers MUST consult revocation metadata supplied with the Audit Pack and MUST reject Compliance Receipts whose signing key was revoked at or before issued_at.

Retention and Long-Term Verifiability The longest retention floor in this profile is 2192 days (six calendar years), set by and ; the EU side has a parallel five-year (1827-day) floor under . Both exceed the typical operational crypto-period of a signing key under recommended key-management practice. Implementations SHOULD use ML-DSA-65 from the algorithm registry () for receipts expected to be verified after the cryptographic lifetime of classical signature schemes ends. Implementations MUST retain public key material for the entire retention window.

Privacy prohibits the inclusion of raw prompts, tool arguments, and credentials in the signed payload. This profile extends that prohibition to the extension fields defined in this document. The risk_class and incident_class values MUST be drawn from controlled vocabularies and MUST NOT carry free-text personal data. Where the underlying Action references a data subject, the payload_digest field MUST cover the data; the data itself MUST be held in a separate store that respects the data subject's rights under applicable law (including but not limited to the General Data Protection Regulation for EU data subjects, the California Consumer Privacy Act and Virginia Consumer Data Protection Act for the corresponding US states, and the HIPAA Privacy Rule where electronic protected health information is involved). A request for erasure that is granted under applicable data protection law MUST be reflected by deletion of the referenced payload, not by deletion of the receipt; the receipt remains as evidence that an Action occurred and was governed by a named policy at a named time.

Anchor Trust The trust assumptions of an anchor depend on the anchor type. timestamp tokens depend on the trust placed in the named Time Stamping Authority. OpenTimestamps commitments depend on the inclusion of the commitment in a public Bitcoin block. A Compliance Verifier SHOULD treat the simultaneous presence of both anchor types as stronger evidence than the presence of only one.

Replay A Compliance Receipt is bound to a single Action via action_ref. Replay of a Compliance Receipt against a different Action is detectable by action_ref mismatch. The 300-second issued_at skew bound stated in limits the window in which a freshly-replayed receipt can be presented as recent. Where the verifier supports it, two receipts sharing action_ref and issuer_id SHOULD be flagged as a candidate duplicate-emission event for human review. This profile does not require verifiers to maintain a cross-receipt index; deployers needing duplicate-emission detection should arrange it at the Audit Pack production layer.

Cross-Regime Conflict Where the same Action is in scope of more than one regime addressed by this document, the producing system MUST satisfy the union of the applicable requirements. Where a SHOULD clause in one regime conflicts with a MUST clause in another, the MUST clause prevails. Where two MUST clauses conflict, the producing system MUST refuse to issue the receipt and MUST log the refusal as a protectmcp:lifecycle Compliance Receipt.

Algorithm Agility This profile inherits its algorithm registry from . Implementations MUST treat the verification of a historical receipt according to the algorithm registry that was in force at issued_at, not the registry in force at the time of verification, provided that the signing key was not revoked.

Issuer-Misrepresentation Residual Per-agent hash chains under detect tampering inside a single issuer's stream but not the cross-agent attack in which a compromised intermediary silently swaps payload bytes between two honest agents. Both per-agent chains validate; action_ref is a correlation anchor, not a cryptographic binding ( Section 2.2). Under -03, a regulator obtains no cryptographic answer to "did the acknowledging agent acknowledge the bytes the originating agent actually sent". This revision introduces counterparty_binding () as the partial mitigation; the following residuals remain.

• Endpoint collusion. If both signing keys are compromised by the same attacker, the attacker produces a coordinated forgery; no signature scheme defends against this case.
• Intermediary holds the originating agent's key. In hosted-agent deployments where the intermediary possesses the originating agent's private key, it can sign anything as either party. Remote attestation of key origin is the appropriate countermeasure and is out of scope here.
• Originator offline at verification time. requires the originating envelope to be retrievable; if unpublished, offline, or rate-limited, the binding becomes unverifiable (liveness loss, observable as failure).
• Fan-out witness gap. When an originator broadcasts to N acknowledgers, each emits an independent pairwise binding; none witnesses any other. Append-only log profiles (future SCITT-style transparency) are deferred to a later revision.
• Key rotation orphan. If the originating agent rotates keys after emission but before an acknowledger binds it, the storage obligation of still requires the old envelope to remain retrievable; if retention discipline fails, the binding orphans.
• Privacy of envelope hashes. envelope_hash is computed over the full signed bytes including A's signature; an observer of B's receipt learns a stable identifier for A's exact action and therefore can correlate B's behaviour across receipts even when A's payload is otherwise confidential. Where this correlation is unacceptable, a commitment scheme (for example, HMAC over the envelope with a per-counterparty key disclosed only to the verifier) is appropriate; this profile does not specify one.
• Real-time prevention. counterparty_binding is detective, not preventive: B has already accepted the bytes by the time the binding is signed. Verifiers detect tampering only at audit time; the in-flight bytes were not blocked. Where prevention is required, transport-level integrity per is the appropriate primitive in addition to (not instead of) this profile.
• Payload-content semantics. The binding proves byte equality, not semantic equality. An intermediary that re-encodes A's bytes into a different but JCS-equivalent canonical form is detected (the SHA-256 differs); an intermediary that swaps A's bytes for entirely different bytes that B's policy happens to interpret as semantically equivalent is detected too. But an intermediary that swaps A's bytes for an A-signed REPLAY of a prior valid envelope from A is not detected by this binding alone; replay protection requires that the verifier also check action_ref and previousReceiptHash uniqueness within the chain segment.

Cross-Agent Integrity Trust Boundary This section is informative. It records operator guidance for cases where channel-level protection is the only available defence and counterparty_binding per has not yet been adopted by both endpoints. For channels between named principals, implementers SHOULD secure the channel using mutually authenticated TLS 1.3 per ; MAY use the tls-exporter channel binding per derived via where higher channel uniqueness is required; and MAY layer HTTP Message Signatures per where intermediaries perform legitimate transformations. Operators MUST NOT interpret transport-layer security alone as evidence of cross-agent byte equality. Only counterparty_binding produces application-layer, signed, replay-after-the-fact evidence answering that question. Topologies where the intermediary terminates TLS (CDN edges, MCP servers, message buses, orchestrators) defeat transport-layer integrity against the threat case of ; in those topologies counterparty_binding is the only defence this profile offers, and the absence of channel-binding evidence in the Audit Pack SHOULD be documented as a known residual.

Compromised Intermediary Between Two Honest Endpoints This section is informative. Where an Action travels from a sending agent A to a receiving agent B through one or more intermediary processes M, and where M is compromised in such a way that M presents byte sequence X to A and a different byte sequence X' to B, neither A's nor B's cryptographic signature detects the divergence in isolation: each endpoint signs the bytes it observed, and each endpoint's per-agent hash chain per remains internally valid. Under draft-marques-asqav-compliance-receipts-03 the only available cross-agent primitive was action_ref as a SHA-256 join key per ; both A's chain and B's chain remained valid in isolation, and divergence was only recoverable through a regulator-driven post-hoc comparison of the two chains. counterparty_binding introduced in closes the case where M silently swaps bytes between two honest endpoints A and B. In this revision (-04), B's acknowledging receipt is REQUIRED to carry an envelope_hash computed over the exact byte stream B received (SHA-256(A's envelope) under the digest-scope rule of ). A verifier resolves receipt_ref to A's stored envelope, recomputes the digest, and compares; a mismatch indicates that the bytes B signed are not the bytes A signed, and the acknowledging receipt MUST be reported non-conformant per . The binding is detective rather than preventive: it does not stop M from performing the swap in flight, but it produces signed, replay-after-the-fact evidence that the swap occurred. The following residuals remain and are not closed by counterparty_binding alone. The list is intentionally honest about the audit-time, not sign-time, nature of the detective evidence: a verifier resolves receipt_ref to A's retained envelope and recomputes the digest at audit time, so any residual reasoning that depends on "A is not in the loop at sign time" is rhetorical, not load-bearing.

• Collusion of M and B. If M and B are jointly compromised, M swaps the bytes in flight and B issues an acknowledging receipt carrying an envelope_hash computed over the altered bytes that B signs as if they were A's. Under counterparty_binding the audit-time verifier resolves receipt_ref to A's retained envelope and recomputes the digest, so the binding is reported non-conformant when A's storage is honest and reachable; M+B collusion alone does NOT silently succeed. M+B collusion silently succeeds only when the collusion ALSO extends to corrupting A's retained envelope, suppressing A's chain segment, or making A's storage unreachable to the auditor; that is, the true residual is M+B+(A's-storage compromise or unavailability). Operator mitigation: anchor A's chain on independent witnesses (combined + anchors per , and OPTIONAL deployer-operated transparency logs) so that A's anchored chain-segment digests are independently recoverable from public evidence; regulator-side comparison of A's anchored chain against B's stored chain detects the divergence even when A's local storage is impeached.
• Collusion of M and A. If M and A are jointly compromised, A signs a fabricated envelope at M's direction and M relays it to B; B verifies M's relay normally, B's counterparty_binding correctly digests the bytes A signed, A's per-agent chain validates, and B's per-agent chain validates. Every cryptographic invariant in this profile holds because the binding correctly attests that the bytes B received were the bytes A signed; the fraud is in A's intent, not in any byte mismatch. This residual is fundamentally outside the receipt model's threat surface: no application-layer cryptographic primitive in this profile distinguishes a fraudulent A-signed envelope from an honest A-signed envelope when M is also colluding to corroborate plausibility (relay logs, timestamping, message ordering). Operator mitigation: separation of duties between issuer (A) and intermediary (M) so that the same operator cannot control both signing keys and relay logs; anchor evidence on independent witnesses under different trust roots so that an attacker controlling A and M still cannot retroactively coordinate anchor inclusion across uncolluding timestamping authorities; out-of-band attestation by the regulator or auditor of A's operational context (provenance, code signing, runtime attestation) where the policy regime authorises it.
• Compromise of B itself. A B that has been compromised (private key extraction, supply-chain compromise, or insider operation) can sign any envelope_hash the attacker chooses; counterparty_binding proves only that the signing key acknowledged some bytes, not that those bytes match what an honest B would have observed.
• Loss of A's stored envelope. counterparty_binding requires the verifier to resolve receipt_ref to A's full signed envelope; if A's chain segment is unavailable (retention discipline failure, key rotation orphan, deliberate withholding), the binding becomes unverifiable and the receipt is reported non-conformant on liveness grounds rather than on byte-equality grounds. An adversary who can arrange A-envelope unavailability and then re-emit colluding bytes can degrade the binding from a byte-equality check to a liveness-loss flag.

Operators concerned about these residuals in the absence of single-point cryptographic defence SHOULD:

• Anchor receipts to multiple independent witnesses. Where both and anchors are present per , a coordinated M-B collusion attack must also induce both timestamping authorities to anchor the colluding bytes within the operator's anchor interval, raising the conjunction-cost of the attack. Operators MAY add further anchors (e.g. a Deployer-operated transparency log or a witness service) without changing the wire format defined here.
• Use side-by-side chain comparison under regulator subpoena. The audit-trail alternative semantics established for SEC 17a-4 recordkeeping (see ) and the post-market surveillance regime of EU AI Act Articles 12 and 26 (see and ) authorise the regulator to compel both A's and B's Audit Packs and to reconstruct the relay by joining on action_ref per . counterparty_binding reduces the regulator's workload from "compare both chains and detect divergence" to "verify B's bound digest against A's stored envelope"; the underlying subpoena-and-compare workflow remains the regulator's ultimate authority and remains operative when the binding is unverifiable.
• Document M's relay logs out-of-band. Where the intermediary M is identifiable (a named MCP server, message bus, orchestrator, or relay), the operator SHOULD require M to produce signed relay logs covering the time window of the Action and SHOULD submit those logs to the same Audit Pack production layer as A's and B's chains. Out-of-band relay logs do not require a wire-format change in this profile; they are operational evidence that complements counterparty_binding rather than replacing it.

The worst-case latency to detection for an M-only compromise (not M-B collusion) is bounded by the anchor interval recommended in : 24 hours by default, one hour for Deployers operating under Article 17, 23 NYCRR 500.17, or . After the next anchor commits, the divergence between A's anchored chain segment and the digest carried in B's counterparty_binding is permanently recoverable from the anchored evidence alone, without trust in M.

IANA Considerations This document requests two new IANA registries to support stable, machine-checkable extensions to the Compliance Receipt format.

Compliance Receipt Extension Fields Registry IANA is requested to create a new registry titled "Compliance Receipt Extension Fields" under a new "Compliance Receipts" registry group. This registry covers both signed-payload fields and envelope-level fields (siblings of payload and signature); each entry's Description identifies which. Each entry contains:

• Field Name: a JSON object key, lowercase ASCII letters, digits, and underscore.
• Description: a one-line summary of the field's purpose.
• Reference: the document that defines the field's semantics.
• Vocabulary: a URL or registry pointer for the controlled vocabulary that field values are drawn from, or "free-form" if none.

The registration policy is Specification Required, per . The Designated Expert(s) SHOULD verify that the field name does not collide with any field defined by , that the Reference is a stable, dereferenceable specification, and that the Vocabulary is documented sufficiently for an independent verifier to validate values. Initial registry contents (each entry labelled with Scope to disambiguate signed-payload fields from envelope-level fields per the registry description above):

• risk_class - Scope: signed-payload - Risk classification term under the Deployer's risk management documentation - This document - Vocabulary referenced in Audit Pack metadata.
• incident_class - Scope: signed-payload - Incident classification term spanning DORA Article 18(1) (with further specification in and the canonical reporting enumeration of Annex II field 3.23 of ), 23 NYCRR 500.1 Cybersecurity Event/Incident, Covered Cyber Incident, and HIPAA security incident under 45 CFR 164.304 - This document - Audit Pack metadata.
• counterparty_binding - Scope: signed-payload - Signed-payload object carrying a base64-encoded SHA-256 digest (envelope_hash) of a peer agent's full signed envelope including signature bytes, a resolvable opaque locator (receipt_ref), an OPTIONAL expected-acknowledger identifier (expect_ack_from), and an OPTIONAL operational transport_label; see for the full member set and the digest-scope rule - This document - Member vocabulary defined in ; digest algorithm is SHA-256 per with base64 encoding per .
• anchors - Scope: envelope-level - Envelope-level array of timestamp / transparency-log anchors covering the signed envelope; entries carry a REQUIRED type discriminator (rfc3161 or opentimestamps) and a REQUIRED value field, plus OPTIONAL informational members status (anchored / pending / failed) and bitcoin_block (string Bitcoin block hash for upgraded OpenTimestamps entries); full schema is defined in - This document - Anchor type vocabulary: rfc3161 per , opentimestamps per .

This document additionally requests that IANA register counterparty_binding as a new claim in the CBOR Web Token (CWT) Claims registry established by , with this document as the reference and the semantics defined in . The requested CWT claim key is TBD by IANA; this document proposes allocation from the IANA First Come First Served range (claim keys greater than or equal to -65536 and less than -256, or greater than or equal to 65536, per Section 9.1) so as not to consume Expert Review or Standards Action space. The JSON-form claim name is the literal string counterparty_binding as registered above in the Compliance Receipt Extension Fields Registry.

Compliance Receipt Type Namespaces Registry IANA is requested to create a new registry titled "Compliance Receipt Type Namespaces" under the same "Compliance Receipts" registry group. Each entry contains:

• Namespace: a colon-separated identifier prefix used as a value of the type field, lowercase ASCII letters, digits, hyphen, underscore, and colon.
• Description: a one-line summary of the receipt category.
• Reference: the document that defines the namespace.

The registration policy is Specification Required, per . The Designated Expert(s) SHOULD verify that the namespace does not collide with any namespace already registered or any namespace reserved by , and that the Reference is a stable specification. Initial registry contents:

• protectmcp:acknowledgment - A receipt emitted by the B-party in a counterparty_binding pair, confirming receipt of the A-party action and carrying a digest of the bound envelope per - This document.
• protectmcp:decision - A receipt recording a policy evaluation outcome (allow, deny, rate_limit) for an MCP-mediated tool call where a policy was actually evaluated; observation is reserved to protectmcp:lifecycle per - This document.
• protectmcp:restraint - A receipt recording the application or release of a restraint on an agent (e.g., quota, rate limit, sandbox tightening) - This document.
• protectmcp:lifecycle - A receipt recording an agent or system lifecycle event (e.g., configuration change, key rotation, oversight review, or a decision=observation record indicating an Action was signed without policy evaluation per ) - This document.
• protectmcp:lifecycle:configuration_change - A receipt recording a configuration change to an agent or producing system, including changes that disable or re-enable receipt generation (see ); a registered sub-namespace under protectmcp:lifecycle that the reference cloud implementation emits today - This document.

Acknowledgements The author thanks Tom Farley for , on which this profile is built. This profile would not exist without the field catalogue and envelope structure that the upstream draft defines. The author also thanks the Asqav community for review of early drafts.

Normative References National Institute of Standards and Technology OpenTimestamps ISO W3C National Institute of Standards and Technology Informative References ISO European Parliament and Council European Parliament and Council European Commission European Commission European Commission European Parliament and Council European Commission European Parliament and Council European Parliament and Council National Institute of Standards and Technology State of Colorado, Seventy-Fourth General Assembly State of Texas, 89th Legislature, Regular Session United States Department of Health and Human Services New York State Department of Financial Services United States Securities and Exchange Commission Effective date January 3, 2023; compliance date for amendments to 17 CFR 240.17a-4 May 3, 2023. United States Congress Public Law 117-103 was enacted on March 15, 2022. Implementing regulations are proceeding under a CISA notice of proposed rulemaking at 89 FR 23644 (April 4, 2024); the final rule is pending publication.

Worked Example (Informative) This appendix illustrates a Compliance Receipt that satisfies the EU AI Act Article 26 binding for a tool invocation by a High-Risk AI System deployed by a Financial Entity. The wire shape applies identically to United States bindings; the only differences are the values placed in issuer_id (LEI, EIN, or CIK depending on the regime) and in the risk_class and incident_class vocabularies referenced in the Audit Pack manifest. Field values are abbreviated for readability and are not cryptographically valid. The example shows a mid-chain receipt; a chain-genesis receipt would carry a previousReceiptHash of 64 zero hex characters per . The worked example below is illustrative. The field values listed are real placeholders an implementer can canonicalize, but the keys are shown in human-readable order rather than JCS-canonical lexicographic order; the JCS-canonical bytes used as input to SHA-256 reorder keys lexicographically (so the on-the-wire byte order for hashing is action_ref, decision, issued_at, issuer_id, iteration_id, payload_digest, policy_digest, previousReceiptHash, reason, risk_class, sandbox_state, tool_name, type, with each nested object's keys also lexicographically sorted, per ). The sig value, the anchors value entries, and the hash hex values are deterministic placeholders chosen for shape rather than cryptographic validity. Implementations MUST NOT replay or trust this example as a real receipt; the example is not signed by any allocated issuer_id, is not anchored against any TSA or OpenTimestamps calendar, and does not chain into any retained predecessor. The above receipt satisfies the Article 26 binding because:

• issuer_id is a 20-character ISO 17442 Legal Entity Identifier (LEI) that resolves through the trust anchor metadata in the Audit Pack to the named Deployer;
• policy_digest resolves to a retained policy artefact;
• sandbox_state is enabled, satisfying the High-Risk system constraint of ;
• previousReceiptHash links the receipt into the chain per ;
• both an anchor and an anchor are present per .

Under the DORA Article 17 binding a Compliance Verifier additionally checks the longest applicable sectoral retention floor (1827 days as the default, per ) and, where present, that incident_class flattens to the canonical six-value vocabulary referenced in . Under the United States bindings of Section 6 the same verifier additionally checks the longest applicable retention floor (2192 days as the default for receipts under or ) and, where the Deployer is a NYDFS Covered Entity, a HIPAA Covered Entity, or a CIRCIA Covered Entity, that incident_class resolves to the applicable canonical category for each in-scope regime.

Change Log

draft-marques-asqav-compliance-receipts-04 Cross-agent integrity revision. The threat case is a compromised intermediary that swaps payload bytes between two honest agents while both per-agent hash chains validate independently.

• digest-scope correction. The chain-link digest scope is the canonical signing-input bytes (the JCS-canonical serialization of the predecessor's signed payload object, the same bytes the predecessor's signature covers), NOT the envelope object that additionally includes the signature or anchors top-level keys. The earlier -04 text that said the digest covers "the entire signed receipt object including the signature field" was a Security Considerations over-reach that did not match the reference implementation and would have required every deployed chain to be re-issued. The corrected scope matches both the reference implementation and the upstream Section 5.7 rule. The security rationale is updated accordingly: the chain binds the signed-over content of the predecessor (recomputable offline from the predecessor's payload alone), and cross-agent envelope-including-signature integrity is delegated to counterparty_binding (), which IS at the envelope-including-signature scope precisely because the peer signature is the load-bearing artefact in that case. is updated to match.
• New informative in Security Considerations acknowledges that the single-linear per-agent chain rule of creates a per-issuer serialization bottleneck: a denial-of-service against the predecessor pointer (database lock, network partition, slow IO) bounds chain throughput. Mitigation: issuers SHOULD use bounded predecessor-lookup timeouts, emit a structured protectmcp:lifecycle audit event with a stable chain_emission_blocked reason code when the timeout fires, and document a chain-head recovery procedure for crashed emitters. Parallel per-issuer throughput beyond a single linear chain remains the existing distinct-issuer_id escape hatch.
• residual list is rewritten for honesty. The M+B collusion residual is downgraded from "M+B alone defeats counterparty_binding" to "M+B+(A's-storage compromise or unavailability) defeats counterparty_binding": the audit-time verifier resolves receipt_ref to A's retained envelope and recomputes the digest, so an honest reachable A's storage defeats M+B collusion alone. A new M+A collusion residual is added: M and A coordinating to produce a fraudulent A-signed envelope is fundamentally outside the receipt model's threat surface because every cryptographic invariant holds; mitigation is separation of duties between issuer and intermediary plus anchor evidence on independent witnesses under different trust roots.
• The worked-example appendix carries an explicit disclaimer that the keys are shown in human-readable order (not JCS-canonical lexicographic order), the sig/value/hash bytes are deterministic placeholders, and implementations MUST NOT replay or trust the example as a real receipt. The disclaimer states the JCS-canonical key ordering that an implementer would derive on the wire.
• envelope_hash rule is tightened on three axes. First, the base64 alphabet is no longer ambiguous: standard base64 per Section 4 on emission, OR base64url per Section 5 where the transport requires URL-safe encoding, and verifiers MUST accept both and normalise before comparison. Second, JSON-framed digest scope is now normative and mandatory-to-implement: the JCS-canonical UTF-8 byte sequence of A's signed envelope JSON object per , where the envelope is the three-key object {"payload", "signature", "anchors"} with anchors OPTIONAL and B forbidden from re-canonicalizing or stripping any of the three keys before computing the digest. Third, cross-framing equivalence is explicitly addressed: JCS-canonical JSON, COSE deterministic encoding, and JWS Compact Serialization with JCS-canonical payload produce different byte sequences from the same semantic payload-and-signature, so envelope_hash is framing-specific; a transcoding intermediary that re-frames A's envelope MUST be treated as a tampering event, and verifiers MUST NOT reframe before recomputing.
• entries now carry an explicit Scope tag (signed-payload for risk_class, incident_class, counterparty_binding; envelope-level for anchors). The CWT claim request for counterparty_binding now proposes allocation from the IANA First Come First Served range of Section 9.1 rather than the unqualified "unassigned integer range" placeholder, so as not to consume Expert Review or Standards Action space.
• The reference is corrected to cite the statutory authority (Public Law 117-103 Division Y, codified at 6 U.S.C. 681 et seq.) rather than CISA's general topic page; the March 15, 2022 enactment date is retained, and the pending NPRM at 89 FR 23644 (April 4, 2024) is named in the reference annotation.
• New normative Section 4.6 () defines the counterparty_binding extension field, an in-payload object carrying a base64-encoded SHA-256 digest (envelope_hash) over the full signed envelope of a peer agent (including the peer's signature bytes), a resolvable opaque receipt_ref, an OPTIONAL expect_ack_from identifier (kid/issuer_id of the expected acknowledger), and an OPTIONAL operational transport_label. Section 4.6.3 () defines the MUST-reject rule when the bound digest does not resolve, the storage obligation on the Audit Pack production layer, and the pairwise default for N-greater-than-2 chains.
• The anchors top-level array schema is now defined inline in : each entry MUST carry type and value (base64-encoded anchor token bytes), with OPTIONAL informational status (anchored / pending / failed) and bitcoin_block (Bitcoin block hash for upgraded OpenTimestamps entries). The IANA Extension Fields Registry entry for anchors is updated to reflect the four-member schema.
• The IANA Type Namespaces Registry initial contents now include the registered sub-namespace protectmcp:lifecycle:configuration_change emitted by the reference cloud implementation for configuration-change receipts under .
• CIRCIA retention floor () is rewritten to measure two years from the submission of the most recently required CIRCIA report per proposed Section 226.13(c) of the CIRCIA NPRM at 89 FR 23644 (April 4, 2024), not from the date of the underlying Action.
• HIPAA retention () is rewritten to cite 45 CFR 164.316(b)(2) with the six-year floor expressed as "six years from the date of creation or the date when last in effect, whichever is later" and the analogy basis for applying that floor to audit-log content explicitly named.
• EU AI Act Article 26(6) retention () is rewritten to express the six-month floor in calendar-arithmetic terms per ; the 184-day day-count figure is retained as an informative anchoring-interval floor, and the previously-endorsed 183-day pick is withdrawn.
• now states explicitly that issuer_id values MUST be bare identifiers without a scheme prefix where the scheme is unambiguous (LEI: 20-character alphanumeric self-identifying form), aligning the spec with the reference cloud emitter under the cloud-wire-conformance change.
• New informative Section 9.11 () documents the residual threat that counterparty_binding partially mitigates (M silently swaps bytes between honest A and B) and names three operational mitigations Operators SHOULD adopt: multiple independent anchor witnesses, regulator-driven side-by-side chain comparison under SEC 17a-4 audit-trail workflow, and out-of-band M relay logs.
• Section 7 () records two manifest-level fields shipped in the reference Audit Pack producer: regime_mapping_disclaimer (when per-receipt regime predicates derive from a producer-side mapping the original signer did not co-sign) and stale_pending (per-entry flag set when anchor evidence remains pending after the bound of ).
• Section 8.3 () replaces the prior one-paragraph clause with five SHOULD-emit per-axis fields: regimes_satisfied, anchor_valid_ots, anchor_valid_rfc3161, policy_digest_resolved, duplicate_emission_candidate.
• New informative Section 9.9 () enumerates the residuals counterparty_binding does not solve. New informative Section 9.10 () records channel-level operator guidance citing , , , and and disclaims transport-layer security as a substitute for application-layer byte equality.
• Section 11 () adds counterparty_binding to the Extension Fields Registry and requests registration as a CWT claim per .
• New normative references: , , , . New informative references for channel-level guidance only: , , , .
• New normative bounds the inputs to which the inherited JCS rule of is applied: IEEE-754 floating-point numbers MUST NOT appear in the digest-covered canonical form (callers SHOULD serialize numerics as JSON strings or as integer-rational pairs in the IEEE-754 safe range), and tool-version-specific semantic equivalence (SQL case folding, path normalization, locale-aware string collation, numeric tolerance, URL percent-encoding choices) is OUT OF SCOPE for the chain layer. The chain answers byte equality for one agent identity at one wall-clock time only; semantic equivalence belongs in the policy_digest artefact and the Audit Pack manifest.
• now states explicitly that each issuer MUST maintain a single linear per-agent chain and MUST serialize concurrent in-agent emission (parallel tool calls, thread-pool fan-out) through a single predecessor pointer in emission order. Parallel sub-chains within one agent identity (a per-receipt chain_id discriminator) are NOT defined by this profile; an issuer that requires parallel sub-chains MUST express each parallel path as a distinct agent identity with its own issuer_id, signing key, and chain rooted at the all-zero genesis value.
• Wire tightenings introduced in this revision (additive at the message level, restrictive at the verifier level): (a) the anchors entry value member is now REQUIRED where the upstream profile leaves it OPTIONAL; (b) issuer chains are normatively single-linear per issuer (no parallel chain_id discriminator), so a -03 producer that ran parallel sub-chains under one issuer_id emits non-conformant -04 receipts; and (c) IEEE-754 floating-point numbers MUST NOT appear in the canonical form covered by SHA-256 digests. A -03 receipt with a missing anchor value, parallel sub-chains, or a digest-covered float is not a conformant -04 receipt. Implementations targeting -04 SHOULD re-emit -03 receipts under -04 emission rules. The signing algorithms, the canonicalization transformation itself (JCS), the anchor types, and the regime bindings of Sections 5 and 6 are unchanged. A -03 verifier remains conformant for receipts that do not carry counterparty_binding; -03 verifiers encountering the field will ignore it per Section 4.2 extension semantics.
• Rationale paragraph of is rewritten to ground the IEEE-754 float ban on documented divergence between mainstream JSON serializers (Python json.dumps, Go encoding/json, Java Jackson) and the ECMA-262 Number-to-String algorithm referenced by Section 3.2.2.3 of , rather than on a misstated claim that JCS itself fails to specify float serialization outside the safe integer range. The Unicode-normalization bullet of is corrected to acknowledge that Section 3.1 of mandates as-is preservation of Unicode strings (no NFC default).
• Security Considerations is corrected to reflect the tiered DORA deadlines: the four-hour initial-notice clock (with 72-hour intermediate-report and one-month final-report bounds) per DORA Article 17 and the RTS in , rather than a flat 72-hour reporting deadline. New informative reference added.
• The Annex II field 3.23 (Type of the incident) citation in and the incident_class initial-registry entry of no longer reproduces the enumeration verbatim; verifiers MUST resolve the canonical values from directly.
• extends the wire decision vocabulary with observation, a fourth value reserved to type protectmcp:lifecycle for receipts emitted when an Action was signed without any policy evaluation. The new value is the regulator-honest alternative to a misleading allow on the "no policy matched" path; emitters MUST refuse to issue a protectmcp:decision receipt that carries observation, and verifiers MUST reject the combination. The vocabulary-namespace registry entry for protectmcp:decision in is updated to reflect that observation is reserved to protectmcp:lifecycle.
• New informative appendix () catalogues five capture topologies an operator can use to emit conformant Compliance Receipts in environments where the originating application code cannot be modified to call the receipt-emitting SDK directly: in_process_sdk, network_proxy, browser_extension, ebpf_observer, mcp_proxy. The appendix is non-normative; the capture_topology attribute it defines is OPTIONAL at the wire layer and, where present, lives in the Audit Pack manifest entry rather than inside the signed payload object, so the topology declaration does not alter signed bytes. The five values form a closed initial vocabulary at this revision; sketches a future "Compliance Receipt Capture Topologies" IANA registry under the "Compliance Receipts" registry group with Specification Required registration policy per for a follow-on revision, and names reserved-value avoidance guidance for early extenders.

draft-marques-asqav-compliance-receipts-03 Multi-jurisdiction consolidation. The European Union profile (formerly the only profile in -02) and the United States profile (formerly the separate draft draft-marques-asqav-us-compliance-receipts-00) are merged into a single document with two regional bindings sections: Section 5 (European Union) and Section 6 (United States). Sections 1 through 4 (Introduction, Conventions, Relationship to upstream, Receipt Field Profile) and Sections 7 through 11 (Audit Pack, Verifier, Security, IANA, Acknowledgements) are shared across both regimes. Conventions terms that differ across regimes are now disambiguated with regime suffixes (Deployer (EU AI Act) vs Deployer (Colorado AI Act); High-Risk AI System (EU AI Act) vs High-Risk AI System (Colorado AI Act)). The incident_class extension field now lists every applicable canonical category in one place: ICT-related incident under with the Annex II reporting enumeration, Cybersecurity Event/Incident under , Covered Cyber Incident under , and security incident under . The issuer_id rule now permits EIN or CIK as alternatives to LEI for US Deployers without an allocated LEI. The Tamper Resistance security consideration extends the one-hour anchoring SHOULD to NYDFS 500.17 and CIRCIA in addition to DORA Article 17. The Privacy security consideration extends to GDPR for EU data subjects and CCPA / VCDPA / HIPAA Privacy Rule for US data subjects. The Worked Example notes that the wire shape applies identically to US bindings, with only the issuer_id identifier and the vocabularies differing. IANA registries are unchanged; the Initial registry contents for incident_class now describe the multi-regime category set. No changes to the wire format, the field profile, the hash chain, the anchoring rules, the Audit Pack contents, or the Verifier checks. Section 4.1.6 (sandbox_state) and Section 4.5 (risk_class) corrected to attribute the EU risk-management documentation requirement to Article 9 of (Provider's risk management system) rather than Article 26, with Article 26(1) cited as the deployer's instructions-for-use obligation that links to the Provider's Article 9 documentation. Section 6.5.3 (nydfs-retention) corrected to a single-tier five-year floor under 23 NYCRR 500.6(b) (verified verbatim against LII Cornell); the prior tiered 5-year/3-year split (claimed against the DFS Second Amendment) was incorrect because the Second Amendment does not amend Section 500.6, leaving the 2017 single-tier text in force. The 1096-day three-year audit-trail floor previously stated for NYDFS is removed.

draft-marques-asqav-compliance-receipts-02 Submission-ready EU-only profile. Wire-shape alignment with upstream (payload/signature/anchors envelope; payload_digest object form; tool_name REQUIRED for protectmcp:decision; issuer_id equals kid). EU AI Act and DORA bindings authored against Official Journal text. Anchor MUST (at least one of RFC 3161 or OpenTimestamps); both RECOMMENDED; 7-day OpenTimestamps upgrade deadline profile-imposed. Six-month AI Act floor expressed as 184 days; DORA-bound default expressed as 1827 days. IANA registries created.

draft-marques-asqav-compliance-receipts-01 Initial wire-shape alignment with upstream and addition of dual-anchor, hash-chain, retention, and DORA classification bindings. Subsequent revisions superseded the specific values introduced here.

draft-marques-asqav-compliance-receipts-00 Initial version. Defines a profile of that binds receipt fields to EU AI Act Article 12, EU AI Act Article 26, and DORA Article 17.

Appendix - Capture Topologies for Compliance Receipt Emission This appendix is informational and non-normative. It catalogues five capture topologies an operator can use to emit conformant Compliance Receipts in environments where the originating application code cannot be modified to call the receipt-emitting SDK directly. The topologies are listed in order of payload fidelity, from highest (in-process SDK, full payload digest) to lowest (eBPF observer, connection metadata only). All five emit receipts that satisfy the wire profile of Sections 3 and 4; they differ in trust boundary, payload coverage, and the operational identity that the receipt binds. The intent is to give an operator vocabulary for declaring which topology produced a given receipt, so that a verifier or auditor can interpret the receipt's evidentiary weight without re-deriving the architecture from out-of-band documentation. An operator MAY declare the producing topology by including a capture_topology attribute in the Audit Pack manifest entry for the receipt. The attribute is informational at the wire layer and OPTIONAL. The vocabulary defined below is closed for the topologies catalogued in this appendix; future revisions or third-party profiles MAY extend it through the future-IANA-registry mechanism noted at the end of this appendix.

In-Process SDK The originating application links the receipt-emitting SDK directly and calls it inline with the action being recorded. This is the baseline pattern and Sections 3 and 4 of this document are written against. The receipt carries a full payload_digest covering the action's request bytes; the SDK has direct access to the application's principal identity, the policy decision, and the request body. Vocabulary value: in_process_sdk. Trust boundary: the application process itself; the SDK runs inside the application's memory space and inherits its principal. Threat-model note: captures the full request and response payloads, the deciding principal, and the policy context; does NOT capture out-of-process side effects or actions taken by sibling processes that do not link the SDK. Reference implementation hint: the Asqav Python and TypeScript SDKs published under the asqav-sdk umbrella (Apache-2.0) implement this pattern; an operator MAY substitute any other conformant implementation.

Network-Layer Egress Proxy A customer-owned reverse proxy or egress gateway (Envoy, NGINX, or an equivalent forward proxy) sits in the network path between the application and the downstream LLM provider. The proxy tees the request to a co-located Compliance Signer process, which receives the request bytes, applies the policy evaluation, and emits a receipt via a signer RPC. A DNS-rewrite on-ramp (CoreDNS rewriting the LLM hostname to the proxy address) or a Server Name Indication (SNI) router (SNIProxy at Layer 4) MAY be used to force application traffic onto the proxy without per-application configuration. Vocabulary value: network_proxy. Trust boundary: the customer's network egress; the proxy and signer run under the customer's operational control, and the receipt is signed by a key the customer's signer holds. Threat-model note: captures the request and response bytes that traverse the proxy and the network-layer principal identity (source IP, mTLS client cert if present); does NOT capture traffic that bypasses the proxy (direct outbound from a non-routed host, DNS-over-HTTPS to a hard-coded resolver, or TLS connections to certificate-pinned endpoints that the customer's enterprise CA cannot inspect). The receipt's issuer_id binds the customer's signer, not the originating application; the application's identity, where captured, appears as an attribute resolved through the Audit Pack manifest. Reference implementation hint: CoreDNS (Apache-2.0) for the DNS on-ramp, SNIProxy by dlundquist (BSD-2-Clause) for the SNI router, and Envoy (Apache-2.0) for the Layer-7 reverse proxy plane; none of these are normative requirements.

Browser Extension A managed-browser Manifest V3 (MV3) extension installed on employee workstations via the enterprise MDM intercepts fetch, XMLHttpRequest, and EventSource requests to a configured list of LLM hostnames. The extension POSTs the intercepted request and response bytes to a Compliance Signer endpoint, which emits the receipt. For LLM hosts that pin their TLS certificates, the customer's enterprise root Certificate Authority (CA) MUST be installed in the browser trust store via MDM so that the extension's content-script interception can observe decrypted bytes. Vocabulary value: browser_extension. Trust boundary: the managed browser process on the employee workstation; the extension runs under the browser's sandbox and the employee's interactive session. Threat-model note: captures the full request and response payload for LLM calls initiated from the browser by the human user, and binds the receipt to the browser's principal identity (the user's enterprise single-sign-on subject, where the extension can read it). Does NOT capture LLM calls made by native desktop applications, server-side daemons, or browsers without the extension installed; does NOT capture traffic in incognito or private-window modes unless the extension is explicitly authorised for those contexts. Reference implementation hint: the open-source Chrome MV3 extension scaffolding published by Google under the chrome-extensions-samples repository (Apache-2.0) is a useful starting point; the customer's signer endpoint is the Asqav signer or any conformant implementation.

eBPF SNI Observer A kernel-level extended Berkeley Packet Filter (eBPF) probe attached to the host's network stack observes outbound TLS ClientHello records. The probe extracts the SNI hostname, the JA3 client fingerprint, the source and destination addresses and ports, and the connection timestamp, and emits a lower-fidelity receipt that binds the employee, the device, the wall-clock time, and the LLM host without observing payload content. Vocabulary value: ebpf_observer. The receipt SHOULD carry an informational payload_capture attribute set to false in the Audit Pack manifest, so that a verifier or auditor does not assume payload-level evidence is recoverable. Trust boundary: the kernel of the host on which the probe runs; the probe operates below the application's user-space process and observes traffic regardless of application configuration. Threat-model note: captures the existence and counterparty of an LLM call (the "did the call happen" evidence class) and the device-and-employee binding through host attestation; does NOT capture request or response bytes, the prompt content, the model parameters, or the decision-relevant context. Useful when payload capture is operationally infeasible (TLS certificate pinning that the enterprise CA cannot defeat, third-party SaaS that egresses outside the customer's proxy plane) but the operator still needs evidence that a regulated LLM interaction occurred. Reference implementation hint: Inspektor Gadget (Apache-2.0) and Tetragon (Apache-2.0) both expose SNI and connection-metadata events from eBPF probes; neither is a normative requirement.

MCP Transparent Proxy A transparent proxy sits in-path between a Model Context Protocol (MCP) client and one or more downstream MCP servers, terminating the client's stdio, Server-Sent Events (SSE), or streamable-HTTP transport on one side and re-establishing the same transport to each downstream server on the other side. The proxy observes every tools/call and resources/read JSON-RPC method invocation, signs a receipt at the moment the call is forwarded, and emits a second acknowledging receipt carrying a counterparty_binding (see ) at the moment the downstream server's response returns. Both sides of the call are therefore bound bilaterally, with the proxy's signing key serving as the integrity anchor for the pair. Vocabulary value: mcp_proxy. The receipt SHOULD carry the invoked MCP method (for example, tools/call or resources/read) as an attribute in the Audit Pack manifest entry, so that a verifier can filter by call type without re-parsing payload bytes. Trust boundary: the proxy process and its signing key; the upstream MCP client and the downstream MCP server are both treated as honest endpoints under the threat model of , with the proxy itself being the named intermediary whose tampering risk counterparty_binding mitigates. Threat-model note: captures the full MCP request and response payload, the method name, and the client and server principal identities visible at the transport boundary; does NOT capture MCP traffic that bypasses the proxy or that uses a transport the proxy does not implement. Reference implementation hint: the Asqav MCP transparent proxy published under the asqav-mcp repository (Apache-2.0) and the upstream FastMCP project (MIT) on which it builds; neither is a normative requirement.

capture_topology Vocabulary and Considerations for a Future IANA Registry The five values defined in this appendix (in_process_sdk, network_proxy, browser_extension, ebpf_observer, mcp_proxy) form the closed initial vocabulary for the capture_topology attribute. The attribute is OPTIONAL at the wire layer and, where present, MUST appear in the Audit Pack manifest entry for the receipt rather than inside the signed payload object, so that the topology declaration is producer-side metadata that does not alter the receipt's signed bytes. A verifier MUST NOT treat the absence of a capture_topology attribute as a non-conformance condition; absence simply means the producer did not declare a topology. This document is an Independent Submission and does not request IANA action for the capture_topology vocabulary at this revision. A future revision MAY request creation of a "Compliance Receipt Capture Topologies" registry under the same "Compliance Receipts" registry group described in , with the registration policy of Specification Required per and an initial set populated from the five values above. Until such a registry exists, implementations that extend the vocabulary SHOULD document the new value in a reference specification and SHOULD avoid colliding with the five reserved values above. The Designated Expert(s) for any future registry SHOULD verify that a candidate value names a distinct topology (a different trust boundary or a materially different payload-fidelity class) rather than a variant of an existing one, and that the candidate's threat-model note states what is captured and what is not, in the form used by the five entries in this appendix.