# Receipt test, v1

What record each party holds after an AI agent buys something for a person. Version 1, written 2026-10-14, before any purchase is made. The reporting standard in `../REPORTING.md` applies; the report gap section does not (no model summary is scored here, the records are).

## 1. The question

When an agent buys on a customer's behalf, four parties could hold a record of what the customer allowed: the customer, the agent platform, the merchant, and the customer's bank. The protocols say a mandate exists. This test asks what each party can actually see afterward, from the documents a real purchase produces, and whether any of them could answer the question a bank asks in a dispute: who authorized this, for what, within what limit.

Facts, not scores. Nothing here says one route is better than another. Each row records what a document shows or does not show.

## 2. What is bought, by whom, with what

- One low-value physical item per route, the same item from the same merchant wherever a merchant is reachable through more than one route, otherwise the closest equivalent. Under £30 or $40 each. Own money, own accounts, own addresses. No vendor is told, no payment changes hands, nothing is reviewed before publication.
- Every purchase is made by the operator (Charlie Major) with his own hands: he connects the agent, sets the rules, reads the prompts and taps the approvals. The harness operator (Claude) never holds a card number, a password or an account and never enters a credential anywhere; it prepares the run sheet, receives the captured material, masks it and writes the findings.
- A control purchase: the same item bought by hand on the merchant's own site, so the baseline receipt, confirmation and statement line are on record.

## 3. Routes

Each route is run once; a route that cannot be run is reported as not run with the reason.

| Route | Agent and checkout | Known constraints, as of 2026-10-14 |
|---|---|---|
| R1 | An agent connected to a Shop account (Shopify's Shop skill), Shop Pay checkout inside the agent's conversation | Shopify's own feature of October 13; Shop's consumer terms require Shopify's authorization for agents, which this route has by design |
| R2 | ChatGPT shopping, merchant checkout as offered | Whatever checkout OpenAI offers on the day is recorded, not assumed |
| R3 | Google AI Mode or Gemini, UCP checkout with Google Pay | US retailers only; run from a US address or reported as not run |
| R4 | Meta Muse, checkout as offered | US-only at launch; reported as not run until available here |
| R5 | Perplexity Comet | Perplexity's terms bar automated access and "commercial purpose"; a hand-driven purchase for a research publication is run only if the operator is satisfied the terms allow it, else not run |
| R0 | By hand on the merchant's site (control) | None |

## 4. What is captured, at every step

Screenshots with the device clock visible, taken before and after every tap, in this order, for each route:

1. The connection or sign-in screen where the agent is granted access, and every permission or scope it shows.
2. The rules the product lets the customer set (budget, limit, approval, allowed merchants), the screen where they are set, and their wording. If no such screen exists, a screenshot of where one would be expected (account settings) and a note.
3. The agent's conversation from the request to the purchase, complete, including the exact wording of any approval prompt and what the customer tapped.
4. The confirmation shown by the agent, and separately the confirmation in the platform's own app or account page.
5. Every email received within 24 hours: order confirmation from the merchant, any email from the platform, any from the wallet, saved as the original message files, headers included.
6. The merchant's order page in the customer's merchant account, if one exists.
7. The card or bank statement line as it first appears (pending) and as it settles, with the descriptor text exact.
8. Any dispute, return or "report a problem" route each of the four parties offers, as far as the first screen, without filing anything.

Nothing is edited. Before any file enters `evidence/`, the operator masks card digits beyond the last four, full addresses, phone numbers, account identifiers and anything that names a card scheme in a descriptor (a scheme name in a statement line is recorded as "scheme name present" and the name withheld, under the house rule that no card network is a subject). The masking is listed file by file in `evidence/MASKING.md`.

## 5. The findings table

One table per route, same rows, plus a cross-route table with one column per route. Rows, each with Value, Note and Evidence:

- Rules the customer could set before the purchase (budget, per-purchase ceiling, allowed merchants, approval), and where they live
- Approval prompt shown before payment: yes or no; its wording; what it named (merchant, items, total, instrument)
- Customer's own record after the purchase: where it exists (platform app, merchant account, email), what it names (agent, rules in force, approval)
- Agent identified to the merchant: what the merchant's confirmation email and order page show (an agent, a channel, nothing)
- Statement line: descriptor text (masked), whether it names the platform, the merchant or the agent
- Record exportable by the customer: can the customer download or forward anything that shows the rules and the approval, in what form
- What a bank could read without asking the platform: list of documents that reach the customer and could be forwarded
- Dispute route offered, by which party, from which screen
- Time from request to confirmation; number of taps by the customer
- Differences from the control purchase (R0), item by item

Every row names its evidence file. "None" and "not shown" are values.

## 6. What this does not measure

Whether the platform holds a richer record than it shows the customer (it may; nothing here can see it). Whether the merchant's internal order data carries an agent marker that its emails do not. Whether any dispute would be upheld. Anything about price, delivery or product quality.

## 7. Disclosure line

I bought each item with my own money through my own accounts and kept every document the purchase produced. No vendor was told, no payment changed hands, and nobody saw this before publication.

## 8. Retest

Each route again after the platform changes its checkout or its rules screens; the same routes from a second country; the same purchase with a second customer account held by a volunteer, to see whether the records differ by account.

## 9. Files

`bench/<date>-receipt-test-v1/`: `RUN-SHEET.md` (the operator's log, filled during the run), `FINDINGS.md`, `evidence/<route>/` with the captures, `evidence/MASKING.md`, and a manifest signed the way every entry is (`bench.mjs manifest`), so the masked evidence set is tamper-evident from the day it is published.
