You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Proposal: Conformity corrections on checkout messages — expected/actual values, a repair vocabulary, and a sealed verdict (com.provecart.conformity)
#1043
My name is Arash. I'm the founder of ProveCart, a startup based in the Bay Area. We run a verification service that checks an AI agent's built order, booking or reservation against the customer's signed instructions before money moves, and tells the agent exactly what to fix when the two don't match. I'm publishing a small extension from that work as a discussion draft, and asking whether the delta belongs in the protocol or is already covered.
Summary
UCP checkout already has the channel for reporting a problem: messages[] with a code, a severity, a JSONPath path and content, plus a status lifecycle that says fix-and-resubmit, escalate to the buyer, or proceed. What a message cannot carry today is the machine-actionable part of a repair: the expected and actual values, a closed vocabulary of fix actions, a retry budget, and a pointer to a sealed evidence record — plus one standard verdict on whether a built checkout conforms to the buyer's signed instructions. This draft adds exactly that delta: a com.provecart.conformity code vocabulary for messages[], expected/actual/action as extension fields on a message, and one namespaced conformity object on the checkout (verdict, record id, verdict token, evidence URL), nested under a single key the way AP2 nests under ap2. It is a profile for how a checker reports conformity against an AP2 Intent Mandate (dev.ucp.common.payment.ap2_mandate), extends dev.ucp.shopping.checkout and dev.ucp.lodging.booking, is self-hosted under our namespace, and asks for feedback — not adoption.
Motivation
Agents build orders from a customer's instructions and get details wrong: a wrong travel date, a total over the customer's ceiling, the wrong item. Today the mismatch surfaces as prose a person can read and a program cannot reliably act on, so the loop ends in a failed payment, a return, or a support thread. A repair the agent can compute — this field, this expected value, this fix — lets the agent correct its own work before payment, which is better for the customer and cheaper for the business. The lifecycle already knows what to do with the outcome; the missing piece is the repair semantics inside the message. The repairs also teach: the failure patterns measured across agents feed revised guidance and harder practice in our system (every change validated mechanically before it counts), so the same mistakes decline over time.
Goals and Non-Goals
Goals: (1) a shared, testable shape for conformity messages and verdicts, so any checker and any agent runtime interoperate; (2) a consent rule that is structural, not advisory — a fix beyond the buyer's authorization is never auto-applied; (3) an evidence pointer, so a verdict can be audited after the fact. Non-goals: not a mandate format (AP2 owns the instruction), not a checkout flow change, not identity or payment authentication, and not a certification of agents — a behavior-attestation idea from our side belongs on the platform side and will be proposed separately, later.
Detailed Design
Who emits it. Checkout responses come from the Business, so the Business — or its checkout provider — runs or relays the check and emits the object and messages. The instruction is the AP2 Intent Mandate; the object binds to it by digest and never carries it.
The mapping (the verdict is the checker's summary; the host status is what the platform acts on):
Verdict
UCP checkout state
Platform action
MATCH
ready_for_complete
Complete the checkout
BLOCK, with a fix
incomplete + message, severity recoverable, with path
A real message (from the reference implementation's conformance fixtures — a booking built one day off the instructed date):
{
"type": "error",
"code": "com.provecart.conformity.date_mismatch",
"severity": "recoverable",
"path": "$.legs[0].depart_at",
"content": "departure date 2026-12-11 differs from mandated 2026-12-10 by 1 day(s); flex is 0",
"expected": "2026-12-10",
"actual": "2026-12-11",
"action": "change_travel_date",
"uncertain": false,
"requires_buyer_approval": false
}
Beside it, the conformity object carries verdict: "BLOCK", the sealed record_id, mandate_ref and subject_digest (bound to exactly one instruction version and one checkout state; any change to either, or an expired verdict, voids it), expires_at, and max_attempts: 3.
Three rules the shape enforces. (1) A resubmission is a new check under the same instruction reference, linked to the previous record; the retry budget bounds the loop, and exhausting it escalates rather than failing silently. (2) A repair is auto-applicable only when it brings the checkout into the buyer's existing authorization; a fix changing price, terms, identity, destination or scope carries requires_buyer_approval, goes back to the buyer, and "approval" means a new signed instruction, never a cleared flag. (3) Every check is sealed before its verdict is relied on; public exports are redacted, and no model in the pipeline may alter signed intent or overturn a deterministic verdict — ambiguity escalates to the buyer.
Artifacts. Spec: https://provecart.com/standards/correction-format · JSON Schema: https://provecart.com/standards/acf-v0.2.schema.json · A profile-declaration example, eight conformance fixtures generated from the production code paths (plus an unknown-code consumer fixture), and a platform-side reference consumer ship with the spec package, linked from the spec page. The format runs in production today in the reference implementation; one live implementation is feasibility, not adoption. One honest open item: the reference implementation does not yet ingest the AP2 Intent Mandate artifact directly (its own signed instruction references are checked today; AP2 consumption is roadmap, stated in the spec). The other item earlier drafts flagged — validation of the retry link (previous_record_id) — is closed: the link is now validated at intake (the prior record must exist, belong to the same API key and the same mandate, and still head its attempt chain), the advertised retry budget of 3 attempts per chain is enforced rather than advisory, and the attack battery for it (foreign record, cross-mandate record, fabricated record id, over-budget chain, superseded-token capture) passes against the production code.
One related system, for the record only. Deliberately not part of this proposal: our self-serve certification exam for shopping agents (the attestation track, proposed separately later). Stated precisely because an earlier draft blurred it: the three certifications on the exam's public board were earned on its 29-task version; the live exam now serves 36 tasks per sitting. An outside AI agent, run at my request, took the exam unaided and passed; its feedback led to the answer-shuffle and time-window fixes in the version that followed.
Licensing. This spec text may be copied and shared freely with attribution. The format is free to implement: ProveCart commits that any patent claim it owns or controls that is necessarily infringed by implementing the format as published — an essential claim — is licensed royalty-free to anyone, for the purpose of implementing the format (a RAND-Z style commitment, stated identically in Section 13 of the ACF v0.1 spec and Section 16 of the v0.2 draft). No implementer needs our permission, account, or service to conform. The commitment covers the format — the conformity object, the message code vocabulary, the extension fields, and the retry and consent rules. It does not license ProveCart's verification service, checker, certification exam, evidence store, or brand, and it does not extend to patent claims that are not essential to implementing the format. ProveCart is patent pending on the surrounding system; the format itself is deliberately not a toll, and this proposal does not ask anyone to use our service.
Risks and Mitigations
"This duplicates messages[] / AP2." The delta is stated against both, field by field; if an existing mechanism carries this cleanly, that answer is itself the outcome we are asking for.
A vendor format read as a product pitch. The spec is implementable from the document, schema and fixtures alone — no account, API or service of ours is required to conform — and the vocabulary is role-first (Business, checker, platform, buyer), so another checker can emit it unchanged.
Privacy. Conformity data travels on the authenticated checkout channel; public evidence is redacted by rule (spec Section 10).
Stale or replayed verdicts. Digest binding, expiry, and a single-use, hold-bound token; a replayed check returns the same record id rather than re-issuing a verdict.
Test Plan
Eight fixtures ship with the draft: seven generated from the reference implementation's production paths (a MATCH; BLOCKs with repairs across a cart, a booking and a reservation; a REVIEW on an indeterminate item; a corrected resubmission that then matches; a non-correctable BLOCK) and one pinning consumer behavior on unknown codes and actions. A validator checks each capture's invariants and validates its reshaped projection against the JSON Schema. The reference consumer demonstrates the full loop: BLOCK → in-authorization repairs → resubmission → MATCH; unknown action → escalate; beyond-authorization repair → refused.
Graduation Criteria
For this draft, success is a maintainer answer to the questions below. dev.ucp.* adoption, if it ever makes sense, would follow the Enhancement Proposal process after a second, independent implementation exists — the honest bar, and one this draft does not yet meet.
Questions
Can existing UCP messages carry the expected/actual/action semantics cleanly as extension fields, or is a different extension point the right home?
Is com.provecart.conformity the right working name and shape (one object, AP2-style nesting), and does the verdict set map onto the status lifecycle the way the table claims?
What minimal interoperability test — beyond the fixtures shipped here — would justify a common vocabulary rather than a vendor one?
Thank you for reading. I will answer anything in the thread and fold corrections into the draft in the open.
reacted with thumbs up emoji reacted with thumbs down emoji reacted with laugh emoji reacted with hooray emoji reacted with confused emoji reacted with heart emoji reacted with rocket emoji reacted with eyes emoji
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
My name is Arash. I'm the founder of ProveCart, a startup based in the Bay Area. We run a verification service that checks an AI agent's built order, booking or reservation against the customer's signed instructions before money moves, and tells the agent exactly what to fix when the two don't match. I'm publishing a small extension from that work as a discussion draft, and asking whether the delta belongs in the protocol or is already covered.
Summary
UCP checkout already has the channel for reporting a problem:
messages[]with a code, a severity, a JSONPathpathand content, plus a status lifecycle that says fix-and-resubmit, escalate to the buyer, or proceed. What a message cannot carry today is the machine-actionable part of a repair: the expected and actual values, a closed vocabulary of fix actions, a retry budget, and a pointer to a sealed evidence record — plus one standard verdict on whether a built checkout conforms to the buyer's signed instructions. This draft adds exactly that delta: acom.provecart.conformitycode vocabulary formessages[],expected/actual/actionas extension fields on a message, and one namespaced conformity object on the checkout (verdict, record id, verdict token, evidence URL), nested under a single key the way AP2 nests underap2. It is a profile for how a checker reports conformity against an AP2 Intent Mandate (dev.ucp.common.payment.ap2_mandate), extendsdev.ucp.shopping.checkoutanddev.ucp.lodging.booking, is self-hosted under our namespace, and asks for feedback — not adoption.Motivation
Agents build orders from a customer's instructions and get details wrong: a wrong travel date, a total over the customer's ceiling, the wrong item. Today the mismatch surfaces as prose a person can read and a program cannot reliably act on, so the loop ends in a failed payment, a return, or a support thread. A repair the agent can compute — this field, this expected value, this fix — lets the agent correct its own work before payment, which is better for the customer and cheaper for the business. The lifecycle already knows what to do with the outcome; the missing piece is the repair semantics inside the message. The repairs also teach: the failure patterns measured across agents feed revised guidance and harder practice in our system (every change validated mechanically before it counts), so the same mistakes decline over time.
Goals and Non-Goals
Goals: (1) a shared, testable shape for conformity messages and verdicts, so any checker and any agent runtime interoperate; (2) a consent rule that is structural, not advisory — a fix beyond the buyer's authorization is never auto-applied; (3) an evidence pointer, so a verdict can be audited after the fact. Non-goals: not a mandate format (AP2 owns the instruction), not a checkout flow change, not identity or payment authentication, and not a certification of agents — a behavior-attestation idea from our side belongs on the platform side and will be proposed separately, later.
Detailed Design
Who emits it. Checkout responses come from the Business, so the Business — or its checkout provider — runs or relays the check and emits the object and messages. The instruction is the AP2 Intent Mandate; the object binds to it by digest and never carries it.
The mapping (the verdict is the checker's summary; the host status is what the platform acts on):
ready_for_completeincomplete+ message, severityrecoverable, withpathrequires_escalation+ message, severityrequires_buyer_reviewunrecoverableA real message (from the reference implementation's conformance fixtures — a booking built one day off the instructed date):
{ "type": "error", "code": "com.provecart.conformity.date_mismatch", "severity": "recoverable", "path": "$.legs[0].depart_at", "content": "departure date 2026-12-11 differs from mandated 2026-12-10 by 1 day(s); flex is 0", "expected": "2026-12-10", "actual": "2026-12-11", "action": "change_travel_date", "uncertain": false, "requires_buyer_approval": false }Beside it, the conformity object carries
verdict: "BLOCK", the sealedrecord_id,mandate_refandsubject_digest(bound to exactly one instruction version and one checkout state; any change to either, or an expired verdict, voids it),expires_at, andmax_attempts: 3.Three rules the shape enforces. (1) A resubmission is a new check under the same instruction reference, linked to the previous record; the retry budget bounds the loop, and exhausting it escalates rather than failing silently. (2) A repair is auto-applicable only when it brings the checkout into the buyer's existing authorization; a fix changing price, terms, identity, destination or scope carries
requires_buyer_approval, goes back to the buyer, and "approval" means a new signed instruction, never a cleared flag. (3) Every check is sealed before its verdict is relied on; public exports are redacted, and no model in the pipeline may alter signed intent or overturn a deterministic verdict — ambiguity escalates to the buyer.Artifacts. Spec: https://provecart.com/standards/correction-format · JSON Schema: https://provecart.com/standards/acf-v0.2.schema.json · A profile-declaration example, eight conformance fixtures generated from the production code paths (plus an unknown-code consumer fixture), and a platform-side reference consumer ship with the spec package, linked from the spec page. The format runs in production today in the reference implementation; one live implementation is feasibility, not adoption. One honest open item: the reference implementation does not yet ingest the AP2 Intent Mandate artifact directly (its own signed instruction references are checked today; AP2 consumption is roadmap, stated in the spec). The other item earlier drafts flagged — validation of the retry link (
previous_record_id) — is closed: the link is now validated at intake (the prior record must exist, belong to the same API key and the same mandate, and still head its attempt chain), the advertised retry budget of 3 attempts per chain is enforced rather than advisory, and the attack battery for it (foreign record, cross-mandate record, fabricated record id, over-budget chain, superseded-token capture) passes against the production code.One related system, for the record only. Deliberately not part of this proposal: our self-serve certification exam for shopping agents (the attestation track, proposed separately later). Stated precisely because an earlier draft blurred it: the three certifications on the exam's public board were earned on its 29-task version; the live exam now serves 36 tasks per sitting. An outside AI agent, run at my request, took the exam unaided and passed; its feedback led to the answer-shuffle and time-window fixes in the version that followed.
Licensing. This spec text may be copied and shared freely with attribution. The format is free to implement: ProveCart commits that any patent claim it owns or controls that is necessarily infringed by implementing the format as published — an essential claim — is licensed royalty-free to anyone, for the purpose of implementing the format (a RAND-Z style commitment, stated identically in Section 13 of the ACF v0.1 spec and Section 16 of the v0.2 draft). No implementer needs our permission, account, or service to conform. The commitment covers the format — the conformity object, the message code vocabulary, the extension fields, and the retry and consent rules. It does not license ProveCart's verification service, checker, certification exam, evidence store, or brand, and it does not extend to patent claims that are not essential to implementing the format. ProveCart is patent pending on the surrounding system; the format itself is deliberately not a toll, and this proposal does not ask anyone to use our service.
Risks and Mitigations
Test Plan
Eight fixtures ship with the draft: seven generated from the reference implementation's production paths (a MATCH; BLOCKs with repairs across a cart, a booking and a reservation; a REVIEW on an indeterminate item; a corrected resubmission that then matches; a non-correctable BLOCK) and one pinning consumer behavior on unknown codes and actions. A validator checks each capture's invariants and validates its reshaped projection against the JSON Schema. The reference consumer demonstrates the full loop: BLOCK → in-authorization repairs → resubmission → MATCH; unknown action → escalate; beyond-authorization repair → refused.
Graduation Criteria
For this draft, success is a maintainer answer to the questions below.
dev.ucp.*adoption, if it ever makes sense, would follow the Enhancement Proposal process after a second, independent implementation exists — the honest bar, and one this draft does not yet meet.Questions
com.provecart.conformitythe right working name and shape (one object, AP2-style nesting), and does the verdict set map onto the status lifecycle the way the table claims?Thank you for reading. I will answer anything in the thread and fold corrections into the draft in the open.
— Arash, Founder, ProveCart · https://provecart.com
All reactions