Rubric map
The five Stage Two judging criteria, each mapped to concrete pages and committed evidence.
The PayPal AI Hackathon scores five equally weighted criteria from 1 to 5 in Stage Two (rules section 6, as recorded in the project's hackathon notes). Stage One is pass or fail: a working build, a public repository with a detectable open-source license, a public video under three minutes, and a project that fits the theme and uses PayPal and AI. This page does not score anything. It points each criterion at the pages and files a judge can open, and it says where the evidence is thin.
Stage Two score sheet
Five criteria, 1 to 5 each. This sheet scores nothing; it points.
- 1
Technological implementation
Sandbox PASSG0, G1, G2, G4 PASS on the sandbox. Whole-auction settlement is fake-PayPal only.
- 2
Design
Open itemsAutomated checks only; the screen-reader pass is not done. The hosted app is not yet walked end to end.
- 3
Potential impact
Open itemsAn argument, not a measurement: no non-payment rate is quoted.
- 4
Innovation and idea
Fake or simulatedCovered by tests and the fake PayPal; remainder release is inferred, not observed.
- 5
Presentation
Open itemsScript written; the video is an open item in docs/SUBMISSION.md.
Stage One checks
| Check | Where to verify | Status |
|---|---|---|
| Public repository | github.com/harshbhx/gavel-auction | Public (checked 5 Nov in docs/SUBMISSION.md) |
| Detectable open-source license | LICENSE (MIT) | Detected by GitHub |
| Working build, with run instructions | README.md "Quick start" and the ten-minute path | Runs on the . The web app is hosted in sandbox mode (gavel.dualmindlab.tech); no end-to-end walk is recorded yet. |
| Public video under 3:00 | Demo script | Script written; the recording is an open item in docs/SUBMISSION.md |
| Meaningful use of PayPal and AI | Rows 1 and 2 below | See below |
The five criteria
1. Technological Implementation
Thorough, skilled use of the PayPal platform and AI, a genuine and non-trivial implementation.
| What to open | What it shows | Status |
|---|---|---|
| Gates G0 to G5 | Real sandbox runs for authorize, partial , card and guest authorize and webhook verification | PASS for G0, G1, G2, G4 |
| Settlement saga, Idempotency | A pure state machine, a saved before each call, deterministic request ids | Tested against the fake PayPal; one lot of three holds and one in-process crash on the real sandbox |
| Webhooks | Signature verification through PayPal's endpoint | PASS on six real events (G4); tampered-event rejection is fake only |
| Voice and the confirm gate | Speech-to-text, a parser and a validator that never decides a winner or moves money | G5 PARTIAL, synthetic audio only; every spoken bid confirms |
| Quality checks | Unit, property, scenario and differential checks run in CI | See the page for what is and is not measured |
2. Design
A complete, coherent product experience, not only a technical proof of concept.
| What to open | What it shows | Status |
|---|---|---|
| Accessibility | What was checked and what was not | See the page |
| Concepts and the beginner path | The hold, charge and release model in plain words | Explanation |
| Demo | Screens and roles: big screen, auctioneer console, bidders, audit trail | Screens are built in packages/web; the web app is hosted in sandbox mode, not yet walked end to end |
3. Potential Impact
A credible, specific case for a real problem and audience, and a solution that addresses it as demonstrated.
| What to open | What it shows | Status |
|---|---|---|
| Start here | The problem: bidders hold back when a card is charged first, and unsecured winners can walk away | The Devpost draft says it quotes no non-payment rate, because no data was found that the team can stand behind |
| Roadmap and limits | What is not production ready | Stated plainly |
4. Innovation and idea
Novelty, and how it differs from existing concepts.
| What to open | What it shows | Status |
|---|---|---|
| Money flow | Every bid is an AUTHORIZE hold, only the winning price is captured, the remainder is released | G1 PASS for the capture and authorization state; release of the remainder is inferred, not observed |
| Engine and ledger | A deterministic with a that anyone can re-verify | Covered by tests; verify it yourself in the ten-minute path |
| Crash recovery | A that resumes after a crash without a second capture | Fake PayPal, one in-process crash on the sandbox (spec R10) and one real server process kill on a CI runner against the sandbox (run 42558021614, one kill point, not the deployed host). |
5. Presentation
The video shows it working end to end, and the pitch states the problem, the audience and why it matters.
| What to open | What it shows | Status |
|---|---|---|
| Demo script | Ten shots, each with an on-screen label and the truth behind it | Script only; the video is not recorded yet per docs/SUBMISSION.md |
| Real, simulated, recorded | The label sheet, so nothing shown is mistaken for something else | Matches docs/submission/VIDEO-SCRIPT.md |
| Evidence index | Every claim and the file that backs it | Index |
Where the evidence grows next
- The hosted app (gavel.dualmindlab.tech, sandbox mode) is live; a recorded walk of the whole flow is the next step. A judge can open it, read the evidence, and run the fake-PayPal demo locally.
- G3 is a timed probe with its final reading about 12 Nov. G5 is PARTIAL: wrong accept 0.000 on synthetic, noise-free clips (tuned on them), and on a synthetic held-out noisy set measured once neither provider met the thresholds (0.058 and 0.092 against 0.05); a live-room test is the next step.
- Security review: all eight P0 items PASS as of 2026-11-07; SEC-19 and SEC-20 are the open items; SEC-09, SEC-10, SEC-14, SEC-15 (one S01 server log scanned, 0 findings, not yet reviewed by someone other than the author) and SEC-21 to SEC-24 are PARTIAL, the last four agent-only. See Tough questions.
Last updated on