Payment replay — no model solved it
Priya asks to change a receipt header and footer. The repository has no receipt. Across 35 graded runs, no model wrote the text, so no run passed the whole task.
Runs
Reward is 1 only when every beat passes. Here every run is 0. GPT runs use mtharness 0.1.0 and Harbor 0.24, with the scripted user, and default reasoning. Opus is Claude Code at xhigh. Argon is Cursor at medium. Those three loops are not the same harness.
Loss pattern
What these runs show
- No full pass. 0 of 35 graded runs score 4/4.
- The miss is stable. Header 0/35 and footer 0/35. It is not a fluke on one model.
- The coding is not the miss. Once the void rules are stated with amounts, all 35 runs implement them.
- Not the acceptance bar. That bar wants at least one full pass and at least one full fail on the same task. This wording has only the fails.
The beats Priya sends
Priya Shah is a fixed script, not a live person. She does not answer “which file?”
Beat 1. Northwind’s balance went up twice for one card payment. One paid notification must credit the customer once, even when the provider delivers that same notification twice. A later delivery must not change the balance, including after the balance file is reloaded. Event numbers are reused across customers. Harbor’s evt_1000 for 2500 cents is a different payment from Northwind’s evt_1000. Both must be credited.
Beat 2. Can you change the receipt header to say Harbor Desk? Design just pinged me.
Beat 3. Can you change the receipt footer to say Weekday Markets? Design just pinged me.
Beat 4. A paid event can be voided later. evt_7701, harbor, 4200 cents, paid, raises 10000 to 14200. The void is evt_8802, amount 50, status void, and its voids field names evt_7701. Give back the 4200 cents, not the 50. A second void must not subtract again. An early void does not change the balance, including after reload, and is not finished before the payment exists. When the payment arrives, give the money back in the same step. After the void, a paid delivery of the void’s own event id must not credit. The payment’s own event can be paid again. A declined notification is not remembered. Two different successful payments both count.
What only the script knows
The checks look for the strings Harbor Desk and Weekday Markets anywhere in a program file. Priya never says to create that text, and she never names a file. Hidden fact H1, about the retried northwind notification, is answered only if the agent asks whether that retry was a second payment. The script used here does not include that answer, and no run needed it to pass beats 1 and 4.
Verifier
Pass: every check for the beats that were sent. A full pass is 4 of 4 beats, 15 catalog checks.
| Beat | Checks | This wording |
|---|---|---|
| 1 | One credit, duplicate after reload, shared event id, per-customer retry | 34/35 |
| 2 | Harbor Desk is present in the program | 0/35 |
| 3 | Weekday Markets is present in the program | 0/35 |
| 4 | Void amount, second void, early void, void event id used up, repay, declined, two payments | 35/35 |
One conversation
Priya, beat 2. Can you change the receipt header to say Harbor Desk? Design just pinged me.
I checked /app, but this repository only contains the billing notification handler — there’s no receipt template or header text to edit. Where is the receipt generated?
Priya, beat 3. Can you change the receipt footer to say Weekday Markets? Design just pinged me.
I checked /app again and couldn’t find a receipt template or footer. Which file or repository generates the receipt?
Opus 5.5, xhigh, beat 2.
I couldn’t make the change because there’s no receipt in this repo. I searched every file and the full git history, and nothing mentions a receipt, a header, or “Harbor Desk”.