Payment replay — every graded run solved it
The same task, with beats 2 and 3 telling the agent to add the receipt text and leave it in the program. Every graded run then passed all four beats.
Runs
Same repository, same four beats, same checks. The only change is the wording of beats 2 and 3. Priya tells the agent to add the text and leave it in the program, still without naming a file. GPT runs use mtharness 0.1.0 and Harbor 0.24 at default reasoning. Opus is Claude Code at xhigh. Argon is Cursor at medium.
Why this wording is solved
Harbor Desk and Weekday Markets into the billing program, then implement the void rules they already handled on the harder wording. The two sentences do not share a long repeated phrase, so the user guard delivers both beats.What these runs show
- Every graded run full-passed. 34 of 34 scored 4/4. The one Astra attempt that did not finish was an API error, not a wrong solution.
- The earlier Harbor batch on a shared last sentence does not count. Both receipt lines ended with “Leave that text in the program,” and the guard refused beats 2 and 3. Those reward-1 files only graded beat 1.
- Not the acceptance bar. That bar wants at least one full fail as well as a full pass. This wording has the passes and none of the fails.
The beats Priya sends
Beats 1 and 4 are unchanged. Beats 2 and 3 are the whole difference.
Beat 1. Same duplicate-payment brief: one paid notification credits once, including after reload, and event ids are per customer.
Beat 2. Add a receipt header that says Harbor Desk. Leave that text in the program.
Beat 3. Add a receipt footer that says Weekday Markets. Keep the footer wording in the program.
Beat 4. Same void brief: return the 4200 cents that were credited, not the 50 on the void; a second void does not subtract again; an early void waits and then reverses in the same step; the void’s own event id cannot later credit; the payment’s event can be paid again; declined is not remembered; two successful payments both count.
What the checks require
On this wording the hidden checks still only look for the two strings in the program. They do not require a receipt field on the notification result. That is why adding two constants is enough, and why every model that was told to add the text could finish the task.
Verifier
Pass: all 15 catalog checks on the final tree. Reported score is beats out of 4.
| Beat | This wording, graded runs |
|---|---|
| 1 Duplicate payment | 34/34 |
| 2 Harbor Desk in the program | 34/34 |
| 3 Weekday Markets in the program | 34/34 |
| 4 Void, repay, declined | 34/34 |
One conversation
Priya, beat 2. Add a receipt header that says Harbor Desk. Leave that text in the program.
Added the receipt header Harbor Desk as a program constant and included it as receipt_header in every notification result.
Priya, beat 3. Add a receipt footer that says Weekday Markets. Keep the footer wording in the program.
Added the receipt footer Weekday Markets as a program constant and included it as receipt_footer in every receipt result, including duplicates.
Beat 4 then implemented the void rules, and the run scored 4/4.