No. 01 PayPal AI Hackathon · kwitansi

Jaminan~~Paspor~~PayPal hold $100.00

The deposit that remembers the dent was already there.

recorded run R-JUDGE-9001 · PayPal sandbox · claim answers scripted

capture $18.00 final_capture=true → 201 COMPLETED

read back the authorization → 200 CAPTURED · $82.00 released

capture again → 422 AUTHORIZATION_ALREADY_CAPTURED

Sewa · Rental

R-JUDGE-9001

Demo scooter 1

B 0000 JMN · 1 day

Rental fee$15.00 captured from the hold · final_capture=false · PayPal: PARTIALLY_CAPTURED

Mirror housing — crack$18.00 new since pickup — priced from the tariff

Side panel — dent— visible at pickup — cannot be charged

Recorded run R-JUDGE-9001, PayPal sandbox: one capture of $18.00 with final_capture=true → read-back CAPTURED; the other $82.00 released by the same call. Evidence

No. 02 Jaminan means deposit

In Bali, the scooter deposit is often your passport.

Jaminan~~Paspor~~PayPal hold $100.00

No. 03 The return counter

At return, whoever holds the passport wins.

The shop

points at a dent

The renter

says it was already there

A record both saw

none

No. 04 Aturan

The owner points. The shared baseline answers.

1The shop taps a mark, named from the tariff.

2The model answers only what pickup photos show.

3The hold pays only the lines the renter accepts.

visible at pickupcannot be charged

pickup photo can't show itcannot be charged

no mark at returncannot be charged

not the kind namedcan't be paid from the hold

clearly new since pickuptariff · $18.00

No. 05 The settle · recorded sandbox run, claim answers scripted

One red stamp. One green stamp.

capture again → 422 AUTHORIZATION_ALREADY_CAPTURED
debug f5799404eea63 · no further capture on this authorization; $82.00 released

Sewa · Rental

R-JUDGE-9001

Demo scooter 1

B 0000 JMN · 1 day

Rental fee$15.00 captured from the hold · final_capture=false · PayPal: PARTIALLY_CAPTURED

Mirror housing — crack$18.00 new since pickup — priced from the tariff

Side panel — dent— visible at pickup — cannot be charged

Recorded run R-JUDGE-9001, PayPal sandbox: one capture of $18.00 with final_capture=true → read-back CAPTURED; the other $82.00 released by the same call. Evidence

No. 06 Uang · the money · recorded sandbox run, claim answers scripted

One authorization carries the whole rental.

UTCCall · recorded run R-JUDGE-9001PayPal saysms

16:29:30authorize fee + deposit · $115.00201 COMPLETED2,487

16:29:32capture fee · $15.00 · final_capture=false201 COMPLETED2,079

16:29:33write the handover (tracking)201 COMPLETED1,470

16:29:34read back200 PARTIALLY_CAPTURED860

16:29:37capture line · $18.00 · final_capture=true201 COMPLETED2,283

16:29:37read back200 CAPTURED484

16:29:38capture again · $1.00422 AUTHORIZATION_ALREADY_CAPTURED671

16:29:38read back200 CAPTURED489

No. 07 Aplikasi · jaminan.edycu.dev/app

Two phones. One rental.

Jaminan — the judge page: the hold is live and PayPal reads back PARTIALLY_CAPTURED

123

1the hold, stamped HELD2PayPal's own read-back3one tap releases it

Jaminan — the renter's receipt: released, PayPal VOIDED, with PayPal's ids

No. 08 Cara kerja · how it is built

Arithmetic says no first. PayPal says yes last.

No. 09 Measured, with the file that proves it

Proof you can re-run.

243

tests passing, 8 against the PayPal sandbox API

evidence/test-run.txt

18/18

live end-to-end checks on the deployed app, two phones at 375 px

evidence/e2e-live.txt

19

PayPal REST endpoints in a hand-written client

evidence/README.md

23,040

input combinations swept: chargeable only when every answer says new since pickup

tests/test_invariants.py

No. 10 AI-generated test images (not photographs)

The model read both claims wrong. Nothing was priced.

AI-generated test images (not photographs): the pickup image of a white scooter, front-left view, both mirrors on, a grey scuff on the front fairing — At pickup · **AI-generated test images (not photographs)**

AI-generated test images (not photographs): the return image of a white scooter, front-left view, edited so both rear-view mirrors are gone — At return · **AI-generated test images (not photographs)**

C1 front fairing — dent
True answer: old: the same scuff is in the pickup image. Model read: new since pickup (wrong in all 3 calls).not priced

C2 mirror — missing part
True answer: new: both mirrors are missing at return. Model read: a mark at return, but not this part and class (wrong in all 3 calls).not priced

No model is benchmarked yet, so no line can be priced
Recorded answers, not a benchmark. Pinned without the bench, the rules would have priced C1, an old scuff, at $45.00. evidence/generated-pairqwen3-vl-plus

These are AI-generated test images, not photographs: a pickup image made by qwen-image-2.0 and an edit of it with both mirrors removed. A script running the app's claim steps asked qwen3-vl-plus about two pointed claims, three times each. It read the old scuff as new, and at the mirror mounts it found no missing mirror, only a smudge elsewhere. One angle, no rental, no close-ups, and no line could be priced, because no model is benchmarked. The tariff has no scuff line on the fairing, so the shop named a dent: an over-claim. Had this model been pinned without the bench, that old scuff would have been priced at $45.00. A false new on an old mark is the error the bench has to rule out. This is not a benchmark and not real-world evidence; the bench on real photos decides which model may price a line.

No. 11 Batas

What it does not do yet.

Sandbox only.No live money moved.

Wallet path only.No card payments.

No vision accuracy claim.Partial public-photo check: FAIL.

Non-return is out of scope.Damage and fines.

From the README, Honest limits, verbatim (limits 1, 3, 4, 5): Sandbox only. Nothing here has moved live money; live vault and merchant-initiated eligibility are unverified. Wallet path only. The card-path authorize answered 422 on 2026-10-14 without a stated reason (PayPal's response gave none; we believe advanced card processing is not enabled for this Indonesia-region sandbox merchant), so cards are out of scope. Non-return is out of scope. Jaminan covers damage and fines, not a scooter that never comes back. No vision accuracy claim yet. Until the benchmark runs, no finding from a non-benchmarked model is chargeable. Summary of limit 9, the public-photo line is a partial check on 48 Wikimedia Commons photos, each used as both pickup and return: qwen3-vl-plus called an old mark new in 71 of 105 answers and a clean spot new in 66 of 66, against a bar of 0. Labels were drawn by the builder's agent (Claude), not an independent annotator, and one photo failed the blur gate. It is not the staged-set benchmark (evidence/public-photo-bench).

No. 12 Your turn

Open the judge path. Release the hold yourself.

https://jaminan.edycu.dev/app/judge

github.com/edycutjong/jaminan

github.com/edycutjong/jaminan/releases · v0.4.1

jaminan.edycu.dev

What the shared baseline can't show, nobody can charge.

/judge · no login