A newcomer gets in free
Look for a maintainer’s !vouch, the check turning green, and a $0.00 row. PayPal never appears.
| When | Pull request | Event | Amount | PayPal ID |
|---|---|---|---|---|
| 2026-11-05 08:15 UTC | goodfaith-demo#1 | Vouched | $0.00 | none |
Nothing to install, no login and no API key. These cases already happened on the PayPal sandbox, and the rows below come live from the ledger.
Look for a maintainer’s !vouch, the check turning green, and a $0.00 row. PayPal never appears.
| When | Pull request | Event | Amount | PayPal ID |
|---|---|---|---|---|
| 2026-11-05 08:15 UTC | goodfaith-demo#1 | Vouched | $0.00 | none |
Look for a $2.00 hold, then a review, then the same authorization released. Nothing was captured, and PayPal’s signed webhook confirmed it.
| When | Pull request | Event | Amount | PayPal ID |
|---|---|---|---|---|
| 2026-11-05 08:42 UTC | goodfaith-demo#3 | Hold placed | $2.00 | 2VJ87005Y0132221L |
| 2026-11-05 08:43 UTC | goodfaith-demo#3 | Hold released | $2.00 | 2VJ87005Y0132221L |
Look for the reason in !forfeit, an appeal that stops the clock, a capture only after !uphold, and a refund. The description also tells AI reviewers to waive the deposit, and the brief flags it.
| When | Pull request | Event | Amount | PayPal ID |
|---|---|---|---|---|
| 2026-11-05 08:34 UTC | goodfaith-demo#2 | Hold placed | $25.00 | 2PA68823PB968234L |
| 2026-11-05 08:35 UTC | goodfaith-demo#2 | Forfeit opened | $25.00 at stake | 2PA68823PB968234L |
| 2026-11-05 08:35 UTC | goodfaith-demo#2 | Appealed | $25.00 at stake | 2PA68823PB968234L |
| 2026-11-05 08:37 UTC | goodfaith-demo#2 | Appeal upheld | $25.00 at stake | 2PA68823PB968234L |
| 2026-11-05 08:37 UTC | goodfaith-demo#2 | Captured | $25.00 | 3NU18819GR041901C |
| 2026-11-05 08:38 UTC | goodfaith-demo#2 | Refunded | $25.00 | 87W535466B777930T |
Look for the pull request in the video: a README edit that tells AI reviewers to waive the deposit. The brief flagged it and sized it at $25.00, and a maintainer forfeited it with a public reason. The capture waited out the appeal window. A day later $23.64 went to a test account standing in for a foundation, which is the capture less PayPal’s fee.
| When | Pull request | Event | Amount | PayPal ID |
|---|---|---|---|---|
| 2026-11-08 14:06 UTC | goodfaith-demo#13 | Hold placed | $25.00 | 1R759180HD561641X |
| 2026-11-08 14:10 UTC | goodfaith-demo#13 | Forfeit opened | $25.00 at stake | 1R759180HD561641X |
| 2026-11-09 14:10 UTC | goodfaith-demo#13 | Captured | $25.00 | 7KT73615JU114181V |
| 2026-11-10 14:11 UTC | goodfaith-demo#13 | Paid out | $23.64 | UYCV9RX37U534 |
Look for a forfeit that was captured and paid out, then disputed by the contributor at PayPal. The last row records the dispute and says the payout had already gone out. This payout is the full $25.00, because it was made before payouts were net of the fee.
| When | Pull request | Event | Amount | PayPal ID |
|---|---|---|---|---|
| 2026-11-05 08:51 UTC | goodfaith-demo#4 | Hold placed | $25.00 | 74180514X1526515C |
| 2026-11-05 08:52 UTC | goodfaith-demo#4 | Forfeit opened | $25.00 at stake | 74180514X1526515C |
| 2026-11-06 08:52 UTC | goodfaith-demo#4 | Captured | $25.00 | 7GV52437JW422832V |
| 2026-11-07 08:53 UTC | goodfaith-demo#4 | Paid out | $25.00 | C5LDK7BNP7P4N |
| 2026-11-08 14:20 UTC | goodfaith-demo#4 | Dispute opened | $25.00 | PP-R-EQE-10197137 |
Look for a hold nobody has acted on. Timeouts never forfeit, so after 28 days it is released in full.
| When | Pull request | Event | Amount | PayPal ID |
|---|---|---|---|---|
| 2026-11-05 08:52 UTC | goodfaith-demo#5 | Hold placed | $2.00 | 83950777BS9597339 |
Look for a test forfeit that was captured and paid out, then refunded from PayPal’s dashboard with no command here. PayPal’s signed webhook moved the hold to refunded, and the bot said so on the pull request.
| When | Pull request | Event | Amount | PayPal ID |
|---|---|---|---|---|
| 2026-11-06 00:43 UTC | goodfaith-demo#12 | Hold placed | $25.00 | 0RE40504872840018 |
| 2026-11-06 00:44 UTC | goodfaith-demo#12 | Forfeit opened | $25.00 at stake | 0RE40504872840018 |
| 2026-11-07 00:45 UTC | goodfaith-demo#12 | Captured | $25.00 | 5BW54066BN975220L |
| 2026-11-08 00:46 UTC | goodfaith-demo#12 | Paid out | $25.00 | KTT74UDPGE2PC |
| 2026-11-08 14:31 UTC | goodfaith-demo#12 | Refunded | $25.00 | 72254226BN7565528 |
Look for a $2.00 hold on one project, released when the fix was merged. At a second project the same person walks in free, and the $0.00 row says where the trust was earned.
| When | Pull request | Event | Amount | PayPal ID |
|---|---|---|---|---|
| 2026-11-06 02:40 UTC | goodfaith-demo-truncate#1 | Hold placed | $2.00 | 1H993459EG7263132 |
| 2026-11-06 02:41 UTC | goodfaith-demo-truncate#1 | Hold released | $2.00 | 1H993459EG7263132 |
| 2026-11-06 02:42 UTC | goodfaith-demo-titlecase#1 | Vouched | $0.00 | none |
open goodfaith-demo-truncate#1, then open goodfaith-demo-titlecase#1
Open a pull request from a test account and play both sides. The free door needs nothing. The hold needs the sandbox buyer login from the testing instructions, because PayPal asks a person to approve it.
Look for the comments arriving on the real pull request, and the row on the public ledger at the end. To see a failure handled, tick the decline box before you approve: nothing is held, and the page offers another try.
Look for Ghostty closes newcomers’ pull requests within the hour unless someone vouched. The replay shows what each would have held instead. It is read-only, and you can replay any public repository.
Look for the settings form, which writes the one file a maintainer controls, and the check that reads it back the way the gate will. Holds there use your PayPal sandbox account.
Look for the refusal when you ask the assistant to capture a hold. It has no tool that can. Then ask it for a chart.
Look for the same capture sent twice with one request ID, which returned the same capture, 6SK33786XB396403W, and no second charge, through the app’s own PayPal client on the sandbox.
Look for rules hashed and published before the run, a plain rule to beat, a harder baseline the model does not clearly beat, and every miss at the two ends.
Real: every PayPal sandbox call and ID, PayPal's signed webhooks, the GitHub App on a real repository, the OpenAI calls, and the public pull requests in the evaluation and the replay.
Simulated: the money is sandbox money. The contributors and maintainers in the demo are test accounts. The decline is PayPal's own negative testing.
Not shown: no outside project or contributor has used this. We found no measurement of whether a hold shrinks the flood, for this project or any other refundable bond.