Bad frames never reach the model
Blur and glare are rejected by arithmetic on a fixed 1280 px frame, before any model call.
numpy + Pillow · Laplacian variance · clipped highlightsJaminanPasporPayPal hold $100.00
At a Bali scooter counter the jaminan, the deposit, is often your passport, and at return whoever holds it wins the argument. Jaminan holds the deposit on PayPal instead: when the shop points at a mark, it cannot be charged if the confirmed pickup baseline already lists that damage there or, once a model passes the bench, the vision model finds it in the pickup photos, and the deposit pays only the lines the renter accepts.
Recorded on the PayPal sandbox · claim answers scripted until the vision bench
final_capture=falsefinal_capture=true releases the rest · or voidPayPal's read‑back after the fee capture and after the final capture or void · all three runs ↓
Recorded replay. PayPal sandbox calls at their recorded speed; the two claim answers are scripted until the vision bench. All three runs ↓
final_capture=false · PayPal: PARTIALLY_CAPTUREDfinal_capture=true → read-back CAPTURED; the other $82.00 released by the same call. EvidenceNo. 1 Aturan · the rule
At pickup, renter and shop walk eight angles round the scooter and confirm the photos together. At return the shop taps each mark it wants to charge and names it from the published tariff. A vision model answers narrow questions about the photos both sides confirmed. It never sets a price.
mirror_housing · crack · $18.00visible_at_pickup · pickup_comparable · mark_at_returncapture · final_capture=trueWhat the shared baseline can't show, nobody can charge.
Code turns the model's answers into one of these lines, in this order. Only the last kind of line has a price, and only from the shop's published tariff.
Who it is for: family-run scooter counters in Jakarta and Bali, and the tourists and locals who rent from them and today leave a passport or KTP as the deposit.
No. 1½ Coba · your turn
AI-generated test images (not photographs) A pickup image made by qwen-image-2.0 and a return image edited from it by wan2.6-image, which removed both mirrors. The builder pointed two claims; a script running the same crop, model-call and decision steps as the app's claim path asked the vision model three times each. One angle, no rental, no close-ups.
The model read both claims wrong, in all 3 calls each.


The model read both claims wrong, in all 3 calls each. Tap a box on the return image for its recorded answer.
These are recorded answers, not a benchmark and not real-world evidence. No rental or payment was involved. Had this model been pinned without the bench, the rules would have priced C1 at the tariff's dent price, $45.00, for a scuff that was there at pickup. A false “new” on an old mark is the error the benchmark has to rule out before any model is pinned, so no model can price a line until it passes the bench on real photos.
No. 2 Uang · the money
One renter approval authorizes the fee and the deposit together ($115.00) and saves the wallet for later fines. Everything after happens on that one authorization, and every stamp waits for PayPal's own read-back.
Rentals longer than three days are reauthorized once on day 4, inside PayPal's day 4 to 29 window. On the sandbox, reauthorize answered 201 on two holds, one with its fee already captured, and refused a second reauthorize of an already reauthorized hold with 422 REAUTHORIZATION_TOO_SOON. If PayPal refuses one, the service authorizes again on the saved wallet and voids the old hold. Inside the app, the day-4 trigger rule and this fallback are unit-tested, including through cron_daily(); the scheduled daily job has not yet run on a rental, so the recorded sandbox calls are the live proof. evidence/paypal-reauthorize.jsonl
Recorded 2026-10-14 16:29 UTC on the PayPal sandbox by scripts/record_paypal_evidence.py: three rentals, each a merchant-initiated authorization on a saved sandbox wallet. The claim answers in these runs are scripted; the PayPal side is real. evidence/paypal-sandbox-run.json
No. 3 Aplikasi · the live app
The same flow runs at jaminan.edycu.dev/app. /judge mints a fresh sandbox rental on a saved sandbox wallet, with no login, and one tap releases it; priced claims wait for the vision bench. Screenshots of the live app; each names its rental.
PARTIALLY_CAPTURED; one tap voids the hold and the green stamp fires on PayPal's read-back, VOIDED.


No. 4 Cara kerja · how it is built
Blur and glare are rejected by arithmetic on a fixed 1280 px frame, before any model call.
numpy + Pillow · Laplacian variance · clipped highlightsPer tap: is there a mark now, is it the kind named, was it visible at pickup, can the pickup photo show that spot.
Qwen VL · JSON schema · 0–1000 boxesThe answers become a line in a fixed order. A model the benchmark has not pinned can never produce a chargeable line.
rules.py · decide_claimPayPal's shipment tracking on the fee capture is the handover record: its tracking number carries the first 8 hex characters of the confirmed baseline's SHA-256.
orders/{id}/trackA traffic-camera ticket whose plate and time match the rental is charged to the saved wallet, capped by the ceiling for that violation; the renter has 72 hours to contest. Offline-tested only: the recorded runs prove a merchant-initiated authorize on a saved wallet, not a fine.
merchant-initiated · vault_idNo. 5 Batas · honest by design
Sandbox only. Nothing here has moved live money; live vault and merchant-initiated eligibility are unverified.— README · Honest limits 1
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.— README · Honest limits 3
Non-return is out of scope. Jaminan covers damage and fines, not a scooter that never comes back.— README · Honest limits 4
No vision accuracy claim yet. Until the benchmark runs, no finding from a non-benchmarked model is chargeable.— README · Honest limits 5
Nothing about money. For each mark the shop taps, the vision model answers narrow questions about the shared pickup photos: is there a mark now, is it the kind named, was it visible at pickup, can the pickup photo show that spot. Code turns the answers into “cannot be charged” or a line priced from the tariff.
Not yet: the benchmark on the staged scooter set has not run. A partial check on 48 public Wikimedia Commons photos, each used as both pickup and return, with labels drawn by the builder's agent (Claude), not an independent annotator, failed: 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. One photo failed the blur gate and was not scored. It does not test new damage between two photos or real shop conditions (evidence/public-photo-bench). Until a model is pinned by it, every live claim ends “comparison unavailable — nothing charged from the hold”, which is why the recorded settlements on this page use scripted answers. The “Tap a claim” section shows the model's recorded answers on AI-generated test images (not photographs); it read both claims wrong, which is a reason for the bench, not a measurement. No benchmark accuracy number appears anywhere; the only real-photo numbers are this failed partial check.
Two locks. The pickup baseline both sides confirmed: a mark it shows, or a spot it cannot show, cannot be charged. And the renter: the hold pays only the lines they accept.
Every contested line leaves as an itemized PayPal invoice at the published tariff price, with a link to the evidence pack, and the hold is released. That is the trade: the hold never pays a line the renter did not accept, and the shop keeps a priced claim on photos both sides confirmed.
Yes, through four MCP tools: list_vehicles, quote_rental, start_rental and get_rental. The agent can price a rental and ask for it, but start_rental returns a PayPal link that only the renter can approve. No tool approves, answers a damage line, captures, voids or refunds. Claude Code and Codex each booked one on the sandbox on 2026-10-24 and read back the authorization; the transcripts are in evidence/mcp.
It mints a fresh sandbox rental on a saved sandbox wallet, with no login, and opens it on the shop's phone; today it can only release the hold (read back VOIDED). A full-path button runs PayPal's real approval page with a sandbox buyer.