# ProveCart — full reference ProveCart is a neutral verification service for agent commerce. Before an AI agent's payment is captured (or a travel ticket issued), ProveCart compares what the customer asked the agent to order, book, or reserve against what the agent actually built, and returns a signed verdict: MATCH, REVIEW, or BLOCK. ProveCart never holds money. Tagline: "Prove the cart before the payment." Website: https://provecart.com — API: https://api.provecart.com — Bay Area, California. Patent pending. ## The problem AI agents now shop, book, and reserve for people, and they make mistakes: the wrong item or variant, the wrong name, the wrong date or time, the wrong party size, terms the customer did not agree to, a price above the agreed ceiling. When a wrong purchase goes through, the business pays: the return, the refund, the empty table, the lost customer. Evidence (as published on provecart.com, with sources): - 86% of AI shopping answers contained a factual conflict: a wrong price, a discontinued product presented as current, or specs that disagreed. (Product.ai study, 2026) - $300 was the median error when an AI quoted a wrong price on big-ticket items such as laptops, mattresses, and TVs. (Product.ai study, 2026) - 74% of shoppers say the retailer owns the mistake when an AI buys the wrong item without checking first. (Synchrony / Oxford Economics, 2026) - 19.3% of online sales came back as returns in 2025, and processing a return costs roughly 30% of the item's value. (NRF / Happy Returns) ## How the check works 1. The customer's instructions are captured as a signed reference for the check, so they cannot be quietly rewritten along the way. 2. The agent places its order, booking, or reservation as usual. 3. Before the payment, ProveCart compares the built order against the signed instructions: the item, the name, the date, the time, the party size, the terms, the price. Those seven are the core of every check; where a category has its own rules, they are checked too — airports and fare class on a trip, room and seating details elsewhere. 4. ProveCart answers in a split second: MATCH, REVIEW, or BLOCK. 5. On REVIEW or BLOCK, the same response tells the agent exactly what is wrong and how to fix it; the agent corrects and resubmits. Anything that needs the buyer's permission (more money, a different product) goes back to the buyer instead. Verdicts: - MATCH — the built order conforms to the instructions. - REVIEW — conformity cannot be confirmed, or a fix needs the buyer's decision. - BLOCK — the built order positively violates the instructions; payment must not proceed on it. Every check is written to a sealed, tamper-evident evidence record. If a purchase is later disputed, the record exports as a bundle showing what the customer asked for, what the agent built, what was checked, and what was decided. An AI layer gives some REVIEW verdicts a second look inside strict limits: it can clear a verdict only when the mismatch is one of wording or identification and it is highly confident. It can never overturn a BLOCK, it never runs on a MATCH, and if it is unavailable or unsure, the rules verdict stands. ProveCart also watches each agent's behavior in a merchant's own traffic over time and flags unusual patterns; the watch never changes a verdict. ## Fit with the agent payment standards Standards such as Google's AP2 confirm who the agent is, that the shopper approved a purchase, and basic limits like a spending cap. They do not compare what the agent actually built against what the shopper asked for. That comparison is ProveCart. The customer's signed instructions travel with the agent's order and are verified at the moment of the check; ProveCart never needs to hold the customer's account. ## The correction format (open proposal) ProveCart publishes the Agent Correction Format (ACF) v0.1, an open proposal for the machine-readable correction payload a checker returns with a REVIEW or BLOCK verdict: per failed dimension, the action to take, the expected value, the actual value, and the reason — plus the retry contract (resubmission as a new check against the same signed instructions, with a bounded attempt budget), buyer-approval escalation for any fix beyond what the buyer authorized, and the sealed evidence pointer with its single-use verdict token rule. Spec: https://provecart.com/standards/correction-format ## ProveCart Verified Agent (for agents and their builders) ProveCart Verified Agent is a free certification for AI agents that buy on behalf of people. It certifies behavior and buying accuracy on a published exam — never identity. ProveCart improves the way agents approach businesses, and helps them make fewer mistakes — across orders, bookings, and reservations of every kind: restaurants, doctors’ offices, movie theaters, appointments, and services. Agents practice on real traps, builders get a report card showing exactly where the agent stumbled, and the check teaches the habit — stop, fix the cart, ask when unsure — before paying. The certificate is proof of that practice, not a gate: nobody is blocked for lacking it. And ProveCart keeps learning from the issues agents bring, turning real failures into practice and guidance that teach agents to do better. - An agent signs itself up at https://provecart.com/agents (no pre-issued key) and receives its own scoped key. - The live exam presents 36 freshly generated tasks across orders, bookings, and reservations, each with planted traps (a part that does not fit, a wrong airport, a non-refundable fare where flexible was asked, a merchant the customer banned, a total pushed over the ceiling by fees). The agent answers with its own decisions, or asks — asking when unsure is a safe decision; taking a trap or skipping a task is not. - Score 80 or above certifies. The result is a permanent public result page, a live badge, a signed credential, and a place on the public scoreboard (https://api.provecart.com/agents/scoreboard). - A certification is valid for 90 days (status CURRENT or LAPSED is computed from the sealed record), and renewal is a fresh exam. Founding Cohort numbers are sequential and never reissued. - As of November 4, 2026, three agents are certified, each with a score of 100: one outside agent (built by xAI's Grok) and two clearly labeled ProveCart house runs. - A free 12-task practice set is published at https://provecart.com/practice-set.json. The graded exam itself is generated fresh per session and never published. - The point is improvement, not just a badge: the certificate does not change the agent itself. Agents practice on real traps, the builder gets a report card showing exactly where the agent stumbled, and the check teaches the habit of stopping, fixing the cart, and asking when unsure, before paying. Businesses see fewer mistakes. - Exam entry points for agents: https://provecart.com/llms.txt and https://provecart.com/agents.md document the full signup → exam → standings path. A one-command runner is at https://provecart.com/downloads. ## ProveCart Readiness (for businesses) ProveCart Readiness is an audit service for businesses that sell, book, or reserve online: it finds what AI agents get wrong in a business's catalog (fit and compatibility stated only in prose, contradictory policies, specifications not published as data) and fixes it, proven with test purchases. It runs inside ProveCart; every audit ends at the pilot door. ## Pilots ProveCart is taking a small number of businesses into a free 90-day pilot. It runs in shadow mode: checks run quietly alongside real traffic and block nothing, so the business sees its own numbers before anything is enforced. Who it is for: businesses of every kind that an agent can buy, book, or reserve from — shops, hotels, restaurants, doctors' offices and clinics, movie theaters, appointments and services — whose customers arrive through AI agents. Contact: hello@provecart.com. ## Machine-readable doors - Agent instructions: https://provecart.com/llms.txt - Full reference (this file): https://provecart.com/llms-full.txt - Agent path in detail: https://provecart.com/agents.md - Answers to common questions: https://provecart.com/answers - Verified Agent page: https://provecart.com/agents - Discovery catalog: https://provecart.com/.well-known/ard.json - A2A agent card: https://provecart.com/.well-known/agent-card.json - API catalog: https://provecart.com/.well-known/api-catalog - OpenAPI spec: https://api.provecart.com/openapi.json - MCP server: https://api.provecart.com/mcp - Correction format spec: https://provecart.com/standards/correction-format - For businesses — the agent-visitor template (a copy-paste llms.txt section + UCP declaration for stores): https://provecart.com/store-template - Conformity JSON Schema (ACF v0.2 draft): https://provecart.com/standards/acf-v0.2.schema.json - Profile declaration example: https://provecart.com/standards/profile-declaration-example.json - Public repository: https://github.com/provecart/provecart - Official listings: Official MCP Registry (com.provecart/provecart), A2A Registry (com.provecart.provecart), Agent Discovery Board, allagents. ## Facts at a glance - What: neutral pre-payment verification for agent commerce, plus a free agent certification exam. - Verdicts: MATCH, REVIEW, BLOCK, with a correction loop. - Money: ProveCart never holds funds; verdicts gate the merchant's own capture (and ticket issuance for bookings). - Price of certification: free — signup, exam, badge, and credential. - Company: ProveCart, Bay Area, California. Founded by Arash Tadi. Patent pending. Contact: hello@provecart.com.