Network Working Group S. A. Gunuganti Internet-Draft Sealgraf Intended status: Informational November 2026 Expires: 19 May 2027 Owner-Approved Agent Spending with Sealed Evidence: The Approval-Evidence Object for Agent Action Protocols draft-gunuganti-agent-approval-evidence-00 Abstract When a software agent acts for a person — in particular when it spends the person's money — the approval at the centre of the flow is, in current agent protocols, a statement by the agent itself. The Personal Agent Protocol's Operations extension states the consequence in its own security considerations: an approval there does not prove that the user saw the terms or approved them; it "works on trust". This document describes the design of a running system that removes the trust step for the approval half of the problem — an owner-set, machine-checkable mandate; a deterministic check performed and recorded before any spendable token exists; approval by the owner on a surface the agent cannot operate or imitate; a single-use decision token bound to the hash of the exact action approved; and a sealed, owner-held receipt that grades its contents (approval PROVEN; the performer's completion report, including payment facts, REPORTED and PINNED) — and specifies an optional approval-evidence object that agent action protocols can carry so that a relying party can verify an approval mechanically instead of trusting the agent's assertion. Status of This Memo This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79. Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/. Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress." This Internet-Draft will expire on 19 May 2027. Copyright Notice Copyright (c) 2026 IETF Trust and the persons identified as the document authors. All rights reserved. This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Revised BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Revised BSD License. Table of Contents 1. Introduction 2. Conventions and Terminology 3. The Approval Problem 4. The Approval-Evidence Object 5. Verification of the Object 6. A Running Implementation (Informative) 6.1. Design Overview 6.2. The Mandate and the Check 6.3. The Approval Surface and the Decision Token 6.4. Sealing and Verification Principles 6.5. Evidence Grades 6.6. Operating Record 7. Relationship to Other Work 8. Open Issues 9. Security Considerations 10. Privacy Considerations 11. IANA Considerations 12. References 12.1. Normative References 12.2. Informative References Author's Address 1. Introduction An automated agent that spends its principal's money creates a question that someone — a merchant, an issuer, an arbitrator, the principal — will eventually ask: what exactly did the owner approve, under what limits, and what happened next? Agent protocols currently answer the middle of that question with an assertion by the agent. In the Personal Agent Protocol (PAP) Operations extension, a Company proposes an action with exact, immutable terms; the user's Personal Agent obtains the user's approval; the Company performs the action at most once. The approval the Company receives is the agent's statement that the user approved. The extension's security considerations (Section 8.1 of Operations version 2, protocol draft 0.3) state the limit directly: "It doesn't prove that the User saw the terms or approved them. The Company can't see the Personal Agent's interface, so an approval works on trust, like the sign-in types. A Company MAY revoke the client_id of a Personal Agent that sends an approval the User didn't give. A Company that needs proof of the User's approval for an action can ask the User to approve on its own website, signed in, the way Direct Sign-In works. This extension doesn't define that step. Approvals the User signs are an open topic." Three properties of this arrangement generalize beyond PAP, which is simply the design that states them honestly: * The attestor is the constrained party: the agent is the actor whose behaviour the approval is meant to constrain, and an approval channel the agent speaks through cannot distinguish a real approval from a fabricated one. * The approval surface is the agent's interface, which neither the relying party nor a later arbitrator can see. * The record belongs to someone other than the principal: the owner, whose money moved, holds no artifact of their own. This document describes a system, in production since 2 November 2026, designed against those three properties, and derives from it a protocol object (Section 4) proposed for The design position taken in this document exists for a reason larger than the disputes the mechanism settles. As agents begin to spend, shop, and commit on a principal's behalf, the default outcome, unless it is designed against, is that the principal quietly ceases to be the owner of their money in any practical sense: the agent acts, the platform attests, and the principal's part is reduced to having once agreed to terms they never saw. The position taken here is the opposite. The owner sets the terms in advance, decides personally above a line they chose, and holds the evidence themself. Delegation does not have to mean surrender: an agent can be given real power to act while the authority over the money, and the proof of how it was used, stays with the person whose money it is. adoption by agent action protocols. The system is Sealgraf, an approval-and-evidence layer for agent spending: it checks, it records, and it never moves money. 2. Conventions and Terminology The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all capitals, as shown here. (In this -00 draft, the normative force attaches to the approval-evidence object in Section 4 and its verification in Section 5; the system description in Section 6 is informative and states what the reference implementation does.) Owner: The person whose money an agent spends; the principal. ("User" in PAP; "principal" in [I-D.izmaylov].) Agent: Software that proposes and performs purchases for the owner. Mandate: The owner's standing permission, expressed as data: a sequence of immutable versions, each carrying the agent it covers, a category allowlist, a per-purchase cap, a rolling period cap, a currency, a validity window, and an ask line. Ask line: The amount above which a purchase requires the owner's personal decision, whatever the standing permission otherwise allows. Action: One proposed purchase: agent, category, amount in integer minor units, currency, counterparty, description. Action hash: The SHA-256 digest of the canonical encoding of an action; the identifier every downstream record cites. Check: The pre-spend evaluation of an action against the current mandate version, performed by a deterministic rule engine. Decision: The check's sealed outcome: allow, deny, or needs-owner-approval, with a stable reason code. Decision token: A single-use, short-lived token minted only after an allowing decision has been recorded, bound to one action hash. Approval: The owner's personal decision on an escalated action, made on the owner's own authenticated surface and sealed as its own statement. Completion: The performer's post-spend report: completed amount, merchant reference, and optionally a payment block. Receipt: The sealed, owner-held bundle joining the cited mandate version, the action, the decision, the approval statement when one was needed, and the completion. Doorbell: A notification (email, push) that carries facts and a link to the approval surface, and never the decision controls. 3. The Approval Problem Standing permissions widen the problem of Section 1. Operations version 1 permitted a confirmation to cite a standing_permission — the user's earlier blanket permission in free text — with no machine-checkable form and no way for the Company to verify that the permission covered the terms. Version 2 withdrew the delegated form: the Personal Agent "MUST NOT approve one on the User's behalf, even when the User said earlier that they would approve something like it." That solves fabrication by removing delegation; but delegation is the normal case for a spending agent. An owner who must personally approve every purchase has a notification service, not an agent. The design position of this document is that per-action approval and standing permission are both necessary, and that both become evidence problems with the same shape: express the permission as machine-checkable data; move the per-action approval off the agent's surface; bind it to the exact terms; record it before anything spendable exists; and give the resulting artifact to the owner. 4. The Approval-Evidence Object The concrete proposal to the community. PAP Operations (and, in shape, any agent action protocol) can define an OPTIONAL approval-evidence object, carried with or fetched after a confirmation, by which an approval service attests a user approval in a form the relying party can verify mechanically. The object states: operation_id, revision the operation the approval covers operation hash hash of the exact terms object, canonically encoded mandate id + version the standing permission the approval was judged under approver the authenticated identity of the approver, the role it acted in (the owner, or a person the owner invited to approve), and its authentication context (method and freshness, as recorded by the surface) outcome approval or refusal decision time when the approver decided terms digest context the mandate version's sealed digest, where limits exist service signature over the canonical object, with the service's keys discoverable from the service's domain Three properties do the work, and a protocol carrying the object SHOULD require all three of an implementation that issues it: * Third-party-verifiable. The object is signed by the approval service, and the keys are discovered from the service's domain — no prior bilateral relationship between the relying party and the service, and no shared secret with the agent. * Independent of the agent. Issuance MUST NOT depend on any assertion by the agent whose spend is being approved — which is precisely the gap Section 8.1 of PAP Operations names. * Bound to the exact terms. The object cites the operation's revision and a hash of the canonical terms object, so terms edited after approval fail verification rather than silently riding the approval. Three further properties, from operation, constrain how the object is produced: * Record before token. The decision exists as a durable record before anything spendable or performable is issued from it. An attestation that can be minted before the decision is recorded attests to nothing. * Refusals are first-class. If approvals carry evidence and refusals do not, an agent (or a service) can silently retry refused actions. The outcome field covers refusal, and a relying party that has seen a recorded refusal of the same terms can treat it as final. * Approval is not performance. The object proves the approval. What happened next — the company's result, or in payments the settlement — is a separate report by the performer, graded separately (Section 6.5), and the two are never merged into one claim. 5. Verification of the Object Verification for the relying party is mechanical: 1. Check the service signature over the canonical object, against keys discovered from the service's domain. 2. Check that the operation identifier and revision match the ones the relying party issued, and have not expired. 3. Check the terms digest against the terms the relying party proposed. 4. Only then perform, under the carrying protocol's existing at-most-once rules. With such an object, an approval stops being a claim the relying party trusts and becomes evidence it checks — the same posture the Operations extension already takes toward the performance result (the Company verifies the action executed exactly once, by operation id). Adoption asks nothing more of the protocol than a place to carry the object and a rule that, when one is present, it is verified rather than trusted. 6. A Running Implementation (Informative) This section describes, at the level the companion paper [SEE-PAPER] uses, the production system from which Section 4's object is derived. It is an existence proof, not a format specification: the implementation's field-level format is documented separately and is not part of this document. 6.1. Design Overview Sealgraf interposes a check between an agent and its spending, without touching the spending itself: 1. The owner sets a mandate (Section 6.2). 2. Before spending, the agent submits the action for a check. Within limits, the decision allows and a decision token is minted. Above the ask line, the owner is notified via a doorbell and decides personally (Section 6.3). Outside limits, the decision denies, with a reason code. 3. The agent claims its token, spends on any rail (a wallet, a card, PayPal — Sealgraf is not in the payment path), and reports completion. 4. The receipt seals, graded as specified in Section 6.5. The owner holds it and can export a bundle that any party verifies offline. The agent reaches the check over MCP (a tool named check_permission) or over REST. Every act in the flow — mandate version, decision, approval, refusal, abort, completion — is appended to a per-owner hash chain and sealed (Section 6.4). 6.2. The Mandate and the Check A mandate is a sequence of immutable versions; every owner change creates a new version, and every decision cites the exact version it was judged under. One version's terms carry: the agent it covers; a category allowlist; a per-purchase cap; a rolling period cap; a currency; a validity window; the ask line; and a status (active, paused, revoked). Amounts are always integer minor units plus a currency code. Mandates are owner-set and never agent-editable; a revoked mandate cannot be revised. A version may also carry a step-up line: the amount at or above which an approval records, in its authentication context, whether a fresh-authentication requirement was in force and satisfied. When unset, the reference implementation derives the line from the ask line. The decision is a deterministic function of (mandate version, period spend so far, proposed action, time): no model is anywhere in the decision path, which is what makes a decision auditable and a receipt meaningful. Rules fire in a fixed order: agent identity; mandate status and validity window; scope (currency, positive amount, category allowlist, per-purchase cap); period cap; and the ask line, which escalates to the owner. Every outcome carries a stable reason code. The counterparty and description are exactly what the approver reads before allowing a spend, so the check boundary screens them for display spoofing (bidirectional-override and control characters), enforced in the engine — the choke point every transport passes through. The decision is persisted to the ledger BEFORE any decision token exists; if persistence fails, no token is issued. Authorizing reserves against the period cap — usage counts completed amounts plus outstanding unexpired reservations, and decisions are serialised per mandate — so two concurrent checks cannot both spend the same headroom, and an approval re-validates the reservation before anything is minted. Decisions that shape the ask channel itself (an exhausted ask budget, a cooling-off period after repeated refusals) are deny decisions sealed like any other: they mint no token and reserve nothing, but they are recorded. The owner's refusal is never shaped; only the agent's future asking is. 6.3. The Approval Surface and the Decision Token An escalated check notifies the owner (email, push) with the facts and a link — a doorbell, never the door. The Allow and Don't-allow controls exist only inside the owner's authenticated session on the approval service's surface. No approval can be given through any channel the agent can originate. This, and not the later signature, is what makes the approval hard to fabricate. The owner may invite other people (a partner, a bookkeeper) to decide escalations. An invited approver decides in their own authenticated session on the same surface, holds no other authority (no change to the mandate, keys, or settings), and can be removed by the owner at any time. Their approval names their own account in the role "approver", in a statement form of its own, so a receipt never presents an invited approver's decision as the owner's; a verifier that does not know that form rejects the receipt as unsupported rather than misreading it. An approval is sealed as its own statement, carrying: the action hash approved; the authenticated principal who approved — recorded as an authenticated principal and role, never a free-text label; the mandate version the approval was judged against, re-validated at approval time against the current version (status, window, scope, period headroom), not merely the version in force when the agent first asked; the approval time; and, for approvals at or above the step-up line, the authentication context (method and freshness, as recorded by the surface). Freshness is recorded, not judged: verifiers display it and apply no verdict to it; whether a recorded freshness is good enough is the reader's policy. Approval mints a decision token: single-use, bound to the action hash, valid for 15 minutes, stored only as a hash. The token is escrowed encrypted for exactly the credential that ran the check, which may claim it after the decision — claim is not consumption — so an agent that was offline when the owner decided can still collect the token, spend, and complete when it returns, with no new check and no broader access. In operation, agents arm a claim poller (polling roughly every 30 seconds after firing an ask) so the owner's tap alone drives claim, payment, and completion. Refusals are first-class. If the owner refuses a purchase from the pending state, the refusal seals; an unchanged re-check of the same action is then denied by name (action_previously_denied), keyed by the action hash and deliberately surviving later mandate revisions: editing the limits does not resurrect a refused purchase. Approval is not the last check. At completion, the engine re-validates the token, its expiry, and the mandate's current status and scope: a pause, revoke, or tightening revision that lands between approval and spending stops the spend. 6.4. Sealing and Verification Principles Every act in the flow is appended to a per-owner hash chain: each entry's hash covers the previous entry's hash and the exact payload bytes; payloads are stored as the bytes that were signed; and every append first verifies the current chain head, so a corrupted head is refused rather than sealed over. Entries are in-toto-shaped statements [INTOTO] carried in DSSE envelopes [DSSE], each envelope dual-signed (Ed25519 and ECDSA P-256; the P-256 key is held in a cloud KMS and is non-exportable), and signing covers the pre-auth encoding of the exact payload bytes — verification never re-serializes a payload to check it. Periodically, the chain is checkpointed: a Merkle tree head over the entries is dual-signed, and the checkpoint digest — the only bytes that leave the operator — can be witnessed externally (in the current deployment, by two independent RFC 3161 timestamp authorities [RFC3161]; OpenTimestamps witnessing is supported by the same machinery). Checkpoints are currently volume-triggered, which Section 8 names as an open issue. The receipt is the owner's artifact: one sealed bundle holding the cited mandate version's terms, the action and its hash, the decision, the approval statement when one was needed, and the completion block. Bundles export into an offline package — exact signed bytes, public keys, chain position, and a Merkle inclusion proof into a dual-signed checkpoint — that a verifier checks with no account, no API key, and no trust in the operator's servers. An approval inside a bundle is PROVEN only when its entry is included in a checkpoint that itself verified (anchored); chain linkage alone is arithmetic a presenter could recompute over a fabricated statement, so an unanchored approval is reported as exactly that — unanchored, unproven — and a bundle whose receipt required approval but carries no approval statement fails verification outright. Approval can never be silently absent from the package that claims it happened. A claim stronger than its proof (for example, an anchor state of "pending" presented as "anchored") is a verification failure, not a warning. The verifier keeps two verdicts separate, and the separation is part of the format: authenticity (the bytes are what the signer sealed and have not mutated) and currency (whether the cited mandate is still in force now — which offline verification cannot check and reports as not checked). The online checker applies the same custody discipline in the other direction: a receipt number alone opens nothing for a non-owner (it is a label anyone may photograph, not a credential); the receipt's check code — known to the owner and whoever they show it to — is the key. Receipt numbers are assigned sequentially within an owner's chain. 6.5. Evidence Grades The record's central discipline is that its contents carry different evidence grades, and the grades travel with the artifacts to every reader, including verifiers: * APPROVAL — PROVEN. The owner approved these exact terms, on a surface the agent could not operate. Proven when the approval statement is anchored (Section 6.4); an unanchored approval is honestly reported as unproven, never upgraded. * TERMS — PROVEN when anchored, by the same rule. * DECISION — PROVEN as the service's sealed record of its own act. Decisions are deterministic given their inputs; the period-cap leg depends on live usage and is not re-judged offline. * COMPLETION — REPORTED AND PINNED. The performer reports what it did — merchant reference and payment block — and the report is sealed byte-identically into the receipt; a later edit is detectable, and the report is attributable. Sealing does not convert a report into a witnessed fact. A completion is sealed even when it reports a charge above the authorized amount: refusing it would erase the only durable record of the overcharge; the comparison belongs at read time. * COUNTERPARTY-SIGNED SETTLEMENT — NOT HELD. No merchant, issuer, or rail signs the reference implementation's completion; where other frameworks obtain counterparty signatures (Section 7.2), their payment leg is stronger, and this design says so. 6.6. Operating Record The system has been in production since 2 November 2026, and the full loop — agent check, owner approval, automatic token claim, purchase on a real rail, sealed receipt — has run end-to-end with real money. On 2 November 2026, an agent checked a USD 10.27 purchase (PayPal rail) that proceeded under the standing mandate and sealed receipt No. 000019; a second purchase the same day (USD 9.72, paid with a Stripe Link one-time virtual Visa) sealed receipt No. 000025, which carries the payment block. The same day, in one production instance, an identical re-check of a purchase the owner had declined was denied by name (action_previously_denied) — a refusal bound a later presentation of the same action, not only in tests. On 3 November 2026, a USD 9.98 purchase sat above the owner's ask line, so the check escalated; the owner approved from their phone; the agent's claim poller collected the decision token automatically; the purchase completed through Stripe Link (charged USD 9.72, as reported) and sealed as receipt No. 000030, publicly checkable by anyone the owner shows its check code to (the code is printed in the companion paper [SEE-PAPER]). Production state as of 7 November 2026: 257 accounts, 216 tenants, 1,868 ledger entries, 103 sealed receipts, 26 agent connections. The caveat belongs in the text: almost all of those accounts are the operator's own test accounts, and the number of external users is zero. The operating evidence therefore supports claims about the mechanism — that the loop runs, that approvals bind, that refusals hold, that receipts verify — and no claims about adoption, usability at scale, or owner behaviour in the wild. 7. Relationship to Other Work The individual pieces of this design are all claimed somewhere, and this document claims none of them. What it offers is a deployed combination — a permission artifact, an enforced check, an approval surface independent of the agent, an owner-facing payment record, and owner-held sealed evidence, in one running loop — and the distinctions that combination makes visible: 7.1. PAP Operations. Closest in spirit. Its exact, immutable terms and proposed standing permissions are the direct ancestors of the mandate (Section 6.2); Sections 3 and 4 are this document's answer to its Section 8: per-action approval on the owner's own surface, standing permission in machine-checkable form, and the approval-evidence object offered back to the protocol. No conformance is claimed: the reference implementation exposes the check through its own MCP tool and API, not through a PAP implementation. 7.2. Agent Payments Protocol (AP2, v0.2, April 2026; FIDO Alliance). Mandates are user-signed on a non-LLM Trusted Surface, and AP2's payment leg is stronger than this design's where it applies — Checkout and Payment Receipts are SIGNED BY THE MERCHANT AND THE PAYMENT PROCESSOR and hash-bound to the mandate, the counterparty-signed proof Section 6.5 marks as not held. AP2 answers payment-proof inside AP2 flows; the approval object here answers approval-proof for any agent on any rail, before the spend rather than at settlement. 7.3. Mastercard Verifiable Intent (v0.1 draft, March 2026; pilots from February 2026). Couples intent cryptographic objects to agentic tokens on network infrastructure, with the issuer's step-up inside the card network — stronger institutional binding inside that network, and, by the draft's own terms, consulted rather than held (pre-transaction). It is Mastercard-rails-only; this design is rail-agnostic and owner-custodied. 7.4. Agent receipts (Asqav; [I-D.marques-asqav]). Signed receipts of agent actions with post-quantum signatures (ML-DSA), including a bilateral pairing of approval and outcome receipts — the closest receipt-layer neighbours. The differences are of subject: Asqav signs ACTIONS for organizations (its terms exclude consumers' personal use) and its approval records "do not pause/resume agent execution" — evidence about flows a control plane governs elsewhere. This design's artifacts are an OWNER's approval and the payment facts around a purchase, and its check executes. 7.5. Signed permission-and-receipt drafts [I-D.izmaylov]. The closest standards-layer neighbour, and the most instructive comparison. The draft's principal signs the permission with a passkey — a stronger identity act than the account session behind the approvals described here — with mechanically checkable limits (amounts, methods, categories, counterparty, cadence), and its records are, in its own framing, "evidence, not enforcement", with an executor role named but its decision procedure expressly unspecified. The designs differ on both axes at once: the signature is the draft's evidence; here the evidence is the approval's provenance — an accountable principal's recorded act, on a surface outside the agent's reach — plus an enforcement point the draft deliberately leaves open. The verifier's questions differ accordingly: the draft reports soundness, signature validity, and within-permission as separate answers; the verification here adds the anchor discipline of Section 6.4. 7.6. x402. A payment protocol whose signed authorization IS the settlement, on-chain: the strongest possible payment proof, for stablecoin payments. It represents no owner permission and issues no owner-facing approval record; the problems are adjacent, not the same. 8. Open Issues * Interception honesty. An agent that never calls the check leaves no receipt, and the reference implementation cannot stop it — the agent is software the owner chose to run. Adoption makes evasion visible rather than impossible: the owner can reconcile "purchases with receipts" against card statements, and a checked agent that goes silent is itself a signal. The realistic mitigations are platform-shaped: wallet and issuer defaults that ask for a decision token, operating-system agent frameworks with the check built in. The design supports those integrations; it does not pretend to be them. * Checkpoint cadence. Section 6.4's volume trigger leaves low-volume owners' approvals unanchored for long stretches — PROVEN only at anchor. The direction under consideration is a signed tree head per sealing event plus a published anchor clock; not yet implemented, and this draft's claims do not include it. * Test asks. The implementation has no test flag. Test checks run as ordinary asks, distinguished only by unique test items, so denial memory never meets production history; an unflagged test counts against the period total while pending, and a refused test is a real refusal in denial memory. * Platform trusted surfaces. AP2's Trusted Surface and the card networks' issuer flows are stronger approval UX than a web surface; this design's surface-independence argument cuts both ways — the owner approves somewhere the AGENT cannot reach, but also somewhere the PLATFORM does not mediate. An approval service whose surface renders inside a platform's trusted UI, without surrendering record custody, is the open design problem. * Owner-signed mandates. A neighbouring design (Section 7.5) has the owner sign the permission with their own key — a stronger identity act than the authenticated session behind the approvals described here. The counterweight is ceremony: key loss, device change, and recovery become owner-facing problems for a non-technical owner population. The trade is unresolved, and this design has not taken it. * Scope. The system judges purchases (category, amount, counterparty, window). Other agent actions — messages sent, files changed, bookings made — share the approval-evidence shape; nothing here claims they are solved. 9. Security Considerations Signing-oracle surface. A sealing service that signs caller-supplied bytes is an oracle; the reference implementation signs only fixed-schema payloads constructed by the sealing path, and its production P-256 key is non-exportable in a cloud KMS. Record before token. Any implementation of this design MUST persist the decision before minting anything spendable (Section 6.2): a crash between decision and record would otherwise create spendable authority with no evidence. Token custody. Decision tokens are single-use, short-lived, hash-stored, and bound to one action hash; the escrow/claim discipline of Section 6.3 ensures an unclaimed token's plaintext dies with its preauthorization's terminal transition. Presenter fabrication. The most dangerous forgery is not against the signed receipt but against the UNSIGNED records around it: a presenter who fabricates an approval statement and recomputes the chain arithmetic over it. Section 6.4's anchor discipline — PROVEN only when a dual-signed checkpoint covers the entry and the inclusion proof recomputes — is the countermeasure, and verifiers MUST NOT soften it. The symmetrical rule closes the hole for receipts: a correctly hashed but unsigned entry is a forgery, however well it links. Anchor overclaim. A state presented stronger than its proof is a verification failure (Section 6.4), in implementations as in bundles: "pending" MUST NOT render as "anchored". Denial memory abuse. Because an owner refusal permanently denies the identical action, an attacker who can induce refusals can block legitimate purchases; ask shaping (Section 6.2) bounds the ask channel, and owners can always re-scope a purchase (different amount, category, or counterparty) to escape a stale refusal. PAN screening. Completions are scanned for full card numbers and refused if found; the payment block has no field a full card number could legitimately occupy. Display spoofing. Counterparty and description strings are attacker-influenced and are what the approver reads; bidirectional and control characters are screened at the engine boundary (Section 6.2), and renderers of sealed artifacts MUST escape rather than interpret carried text. Key compromise. Key-registry status transitions (active / verify-only / compromised) bound the blast radius: a compromised key's signatures MUST NOT be relied upon, and the dual-signature policy means one algorithm's compromise alone does not authenticate an envelope. 10. Privacy Considerations De-identified approvers. Approval statements name the approver by authenticated principal and role only: never a personal name or email address, so a receipt shown to a merchant or arbitrator does not carry the owner's identity beyond the role that acted. Hash-only anchoring. Anchors disclose exactly one SHA-256 digest per checkpoint to third-party witnesses (Section 6.4); no terms, amounts, or identities leave the operator for anchoring. Receipt numbers are not credentials. A receipt number is a chain position, and checkers MUST treat possession of the check code — not the number — as the capability (Section 6.4). Implementations of number-based lookup MUST NOT leak existence across owners (uniform not-found responses). Purchase data sensitivity. A receipt records categories, amounts, counterparties, and instrument last-fours — financial facts about the owner. Custody is the mitigation: the artifact is the owner's, held for the owner, exportable by the owner, and shared at the owner's direction. 11. IANA Considerations This document has no IANA actions. 12. References 12.1. Normative References [RFC2119] Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", BCP 14, RFC 2119, DOI 10.17487/RFC2119, March 1997. [RFC8174] Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC 2119 Key Words", BCP 14, RFC 8174, DOI 10.17487/RFC8174, May 2017. 12.2. Informative References [RFC3161] Adams, C., Cain, P., Pinkas, D., and R. Zuccherato, "Internet X.509 Public Key Infrastructure Time-Stamp Protocol (TSP)", RFC 3161, DOI 10.17487/RFC3161, August 2001. [DSSE] "Dead Simple Signing Envelope (DSSE) protocol v1.0" — Pre-Auth-Encoding as used in Section 6.4. [INTOTO] "in-toto Attestation Framework — Statement and Predicate model": statements of the form used in Section 6.4. [PAP-OPS] Salazar, E., "Personal Agent Protocol — Operations extension (Purchase and Subscription Operations)", protocol draft 0.3, Operations version 1 (standing permission attestations) / version 2 (Section 8.1 quoted in Sections 1 and 3), https://personalagentprotocol.org/docs/0.3/ extensions/operations . [PAP-ISSUE61] "Agent-side shopping mandates (industry alignment thread)", Personal Agent Protocol issue #61, 2026-10-24 (author summary and thread as read November 2026). [I-D.izmaylov] Izmaylov, P., "Signed Permissions and Action Receipts for Automated Agents", draft-izmaylov-agent-permission-receipts-02, 31 October 2026. [I-D.marques-asqav] Marques, J. A. Gomes, "Compliance Profile of Signed Action Receipts for AI Agents", draft-marques-asqav-compliance-receipts-09, 21 September 2026, and Asqav product documentation, as reviewed November 2026. [AP2] "Agent Payments Protocol (AP2), version 0.2", April 2026; FIDO Alliance adoption; Trusted Surface; merchant-signed Checkout Receipt and processor-signed Payment Receipt. [MC-VI] Mastercard, "Verifiable Intent, version 0.1 draft", March 2026 (pilots from February 2026); pilot communications, 2026. [X402] "x402 protocol" — HTTP 402 payment protocol; signed authorization as settlement (stablecoin rails). [SEE-PAPER] Gunuganti, S. A., "Approval You Can Verify: Owner-Approved Agent Spending with Sealed Evidence", November 2026 (companion paper; canonical text lodged with the author). Author's Address Sai Arun Gunuganti Sealgraf Email: support@sealgraf.com URI: https://sealgraf.com