We reviewed seven agent and commerce specifications to see how they handle eligibility at the moment of action, and what evidence survives afterward. Each one dates the payment or the authorization. One requires eligibility to be settled before checkout completes. None defines a signed record, portable with the transaction, that says the buyer still qualified when the agent acted. The regulators asking about agents have not named that record either. This post sets out how widespread the gap is and what we are doing to close it.

The Facts That Change After Signup

A shareholder gets a discount because she holds the stock. An investor may buy because he is accredited. A wallet may trade because it is on a venue's approved list. An agent may spend because its owner delegated the authority. A counterparty may be paid because it is not on a sanctions list.

Each of these was true on the day somebody checked. Each can stop being true the next day. Shares get sold. Accreditation lapses. A customer revokes the agent. A sanctions list adds a name.

Agents multiply the number of actions that can occur before anyone discovers eligibility has changed. A person transacts a few times a day. An agent can transact thousands of times an hour and never stops to ask. A list that was right at onboarding is a stale list by the time the agent uses it.

What Seven Specifications Define

We read the current text of each specification on November 18, 2026. Each name in the table links to the passage quoted, at the version we read. The question for each: when a fact about the buyer or the agent comes from outside the protocol, is it checked fresh at the moment of action, and does a signed record of that check travel with the transaction?

SpecificationFreshness it definesFacts about the partySigned evidence it defines
UCPRequest signatures carry created and expires timesEligibility claims are "claims, not verified facts" and must be resolved before completion. "UCP does not prescribe how verification occurs."The eligibility outcome is reported as a message code
MPPPayment challenges expireThe Payment scheme "does not identify the client, the payer's principal, or any other party"A receipt for the payment
A2ACaching guidance for the Agent Card"Does not define the scope, representation, validity, or revocation semantics of the authorization decision or credential"Signed Agent Cards
AP2Mandates expireConstraints written into the mandateSigned checkout and payment receipts. Their use for "dispute resolution, retention, and retrieval requirements" is "outside the scope of this specification."
ACPThe delegated payment token expiresThe discount extension lists "coupon validation rules" as out of scopeNone in the documents we read
x402Payment authorizations carry a validity windowNot addressedAn optional signed receipt "confirming that payment was received and service was delivered"
Verifiable IntentEach layer of the delegation expiresVersion 0.1 "does not define a credential-layer revocation protocol"The signed delegation itself

That is not a criticism. It is scope. A payment protocol should carry payment, and Verifiable Intent, the specification Mastercard introduced, does record what the user delegated. Read down the last column, though. The signed evidence says money moved, goods shipped, or authority was granted at some earlier time. In none of the seven does a signed record state that a condition about the party was met at the moment of action. UCP requires the check and leaves the method open. MPP, in the same passage, says a resource may require "an attestation carried in an application-defined header field," which it "does not define, register, or constrain." Both leave a named place for the record.

What the Regulators Have Asked

Regulators see the problem. Federal Reserve Governor Christopher Waller, speaking on September 29: "The question shifts from proving that a buyer is an authorized payer to proving that an agent has the authority to pay on the buyer's behalf." CFTC Chairman Michael Selig, opening the agency's forum on October 28: "How can regulators know who the agent, or suite of agents, is trading on behalf of?" FINRA's 2026 oversight report lists the risk that "agents may act beyond the user's actual or intended scope and authority." The Financial Stability Board's consultation on AI notes "the impracticality of real-time human monitoring of agent decisions as their use scales."

These are questions. None of them says what a check at the moment of action should produce.

The clearest statement of the timing problem is older and comes from insurance. OFAC's FAQ 65 says: "Screening only at the point of policy issuance may expose insurers to sanctions risk, for example, providing financial benefit to subsequently blocked persons."

In tokenized securities the point is contested. In the comment file on the SEC's transfer agent proposal, KoreInside, filing with its transfer agent, wrote that "a previously verified wallet or whitelist entry is not permanent proof of eligibility." Several large firms argued the reverse: verify a wallet once when it joins the list, then rely on the list. We filed on the side of checking at the transfer, and explained why.

So the gap is not confined to one field. It sits under a store discount, a securities trade and an agent's payment. The condition differs. The missing record is the same.

Who Already Checks at the Moment

Some do, and they deserve the credit. Coinbase says its x402 facilitator screens every payment against sanctions lists before it settles. Stripe's shared payment tokens can be revoked, and a revoked token cannot pay. ERC-3643, a token standard for regulated securities, checks the receiving wallet against an identity registry inside every transfer.

These are fresh checks. What differs is what is left afterward. The facilitator's and the processor's answers sit in their own systems. The ERC-3643 check runs on the chain, so the evidence exists, but showing it later means reconstructing what a registry contract held at a past block. A merchant, a venue's examiner, or a customer in a dispute two years on wants something simpler: a compact signed result they can file beside the order and verify without going back to anyone.

The Record Has to Outlive the Signature

There is a second clock. A CFTC registrant keeps most records for at least five years. A broker-dealer keeps account records for six years after the account closes. NIST's draft transition plan (NIST IR 8547, initial public draft) lists ECDSA, the signature family used in the AP2 and Verifiable Intent examples, as disallowed after 2035. A record signed in 2030 has to be trusted past that date.

We searched the comment letters posted on three SEC files as of November 12, 2026: File Nos. 4-927, S7-2026-27 and S7-2026-30, 558 letters in all. Nine scanned letters on the third file could not be searched. We found no other letter explicitly connecting record-retention periods with the signature transition. We covered the arithmetic in Tokenized Stock Records and Quantum Risk.

What the Missing Record Looks Like

The missing piece is small and specific: a signed statement that a named condition was met by a named wallet at a named block and time, which anyone can verify later without calling the party that made it.

That is what InsumerAPI returns. The primitive is wallet auth: read wallet state, evaluate conditions, return a signed boolean. Read, evaluate, sign, keep. The business sets the condition. POST /v1/attest reads the chain at that moment across 37 chains and answers yes or no. Boolean, not balance. The result is signed twice: ES256 and a post-quantum ML-DSA-65 signature. It is good for an access decision for up to 30 minutes, and good as evidence for as long as the holder keeps it. The open-source insumer-verify library checks it offline against the published keys.

The limits matter as much as the claim. InsumerAPI reads blockchains. It does not screen sanctions lists and it does not decide who is accredited. It can answer those questions only where somebody with the authority to decide has recorded the answer on a blockchain: an accreditation issued as a token to the investor's wallet, an account attestation from an exchange, a venue's permission issued as a token, a delegation the customer signed. Where the fact lives on a blockchain, removing it is one transaction, and the next check returns no. Where it does not, this is the wrong tool.

Three cases from our own work show the same shape. A shareholder's wallet holds the tokenized share, the merchant's condition is "holds at least one," and the discount applies only while that is true. A venue checks its permission condition at each trade and files the signed result, the case we set out for venues. An agent's delegation is checked before it pays, the subject of What x402 Doesn't Tell Your Agent. The condition differs each time. The record has the same form.

The protocols do not need to change to carry it. x402 moves the money. InsumerAPI checks the conditions.

What We Are Doing to Close It

The check is live today. The rest of the work is getting the record a place in the rules and in the standards. A filed comment means it was received, and an open proposal means it is in review. Neither means anyone agreed.

By letter. We have put the argument to the SEC on the tokenized stock venue exemption (File No. 4-927), on proposed Regulation Crypto Assets (S7-2026-27), on the proposed Transfer Agent Rules (S7-2026-30) and to its Crypto Task Force. We sent written input to the CFTC Innovation Task Force (our own copy, since the CFTC has not posted it). At NIST's NCCoE we commented twice: on the AI agent identity and authorization concept paper in March, and in October on the DevSecOps project, where we asked that each agent action be decided by a condition checked when the agent acts. NCCoE has not published either set of comments in full, so the first link is our own account and the second is the project page. Each one makes the same request in its own setting: check when the action happens, and keep a signed record that stays verifiable for as long as it must be kept.

In standards work. We started where we hit the problem while building. For the commerce and agent protocols we proposed the slot: UCP pull request 415 for a signed wallet attestation beside OAuth, Verifiable Intent pull request 22 so a mandate can require one before an agent executes, and the A2A trust signals proposal. All three are open and in review.

In the draft Ethereum ERCs, review comments have become text that other authors wrote into their own drafts. ERC-8426 requires a fresh ownership read at request time before a pass can trigger an action. ERC-8481 has an auditor check who was authorized to sign when a report was anchored. ERC-8478 lowered the longest a delegation can run from 365 days to 90. Each revision credits the review.

Every item, with its date and status, is on the standards page.

What to Do Next

If you build agents, check the condition before the agent acts and keep the signed answer with the payment record. A free key is at /developers/quickstart/, and an agent with no account can pay per call over x402 at /pricing/#x402. If you run a venue or issue a tokenized security, decide whether your check runs once at signup or at each action, and what you will hand an examiner in year five. If you run a store and want to recognize shareholders or members at the counter, Skye Meta builds on InsumerAPI and runs that side at skyemeta.com.

Condition-based access across 37 chains

InsumerAPI: evaluate wallet conditions, get a signed result. No secrets. No identity. Free tier available.

View API Docs

Get every post by email

The Inevitable series and builder notes on condition-based access. Free, one or two posts a week.