Gates G0 to G5
Each sandbox gate, its result and the caveat that goes with it.
A gate is closed only when its status is PASS or FAIL with a recorded decision. A gate without evidence reads NOT RUN. This table is generated from docs/GATES.md when the site is built.
| Gate | Question | Result | Run |
|---|---|---|---|
| G0 | sandbox app and buyer accounts | PASS | 2026-11-04 by harshbhx (workflow gate-g0, run 41815593520, branch gate/g0) |
| G1 | partial capture and remainder release | PASS | 2026-11-04 by harshbhx (wallet); 2026-11-05 by harshbhx (card, run G1-20261105T123116Z) |
| G2 | card and guest AUTHORIZE | PASS | 2026-11-05 by harshbhx (owner laptop, run G2-20261105T115228Z) |
| G3 | honor period versus validity and reauthorize | IN PROGRESS | Real probe. Day 0 was 7 Nov 2026 03:06 UTC. Final reading about 12 Nov 06:17 UTC. |
| G4 | real webhook signature verification | PASS | 2026-11-05 by harshbhx (workflow gate-g4, run 41947655945) |
| G5 | STT accuracy on noisy bids | PARTIAL | 2026-11-07, cached eval over recorded provider outputs (PR #1940, issue #1864; second clip set PR VOI-01; error analysis, parser fixes and option trials PR VOI-02) |
G0: sandbox app and buyer accounts
G0 PASSThe sandbox app authenticates (HTTP 200, payments scope) and can create an AUTHORIZE order (HTTP 201 CREATED). Run 4 Nov 2026.
G1: partial capture and remainder release
G1 PASSAfter a final capture the authorization closes as CAPTURED and the order COMPLETED, and a later void is refused with PREVIOUSLY_CAPTURED. Run for a wallet buyer on 4 Nov and a card-funded PayPal account on 5 Nov 2026.
G2: card and guest AUTHORIZE
G2 PASSA wallet buyer and a guest with a sandbox card were both authorized through the PayPal JS SDK v6 (5 Nov 2026). Authorize only: no capture and no void of a card authorization ran.
G3: honor period versus validity and reauthorize
G3 IN PROGRESSThe planned fallback if the window is short: capture within the verified validity window and reauthorize before the hammer. That is the plan, not a result.
G4: real webhook signature verification
G4 PASSSix real deliveries, including a capture and a void, each verified SUCCESS through PayPal's verification endpoint with the raw body spliced into the request (5 Nov 2026). Rejecting a tampered event is covered against the fake only.
G5: speech accuracy on noisy bids
G5 PARTIAL| Clip set | Provider | n | Exact | Wrong accept (limit 0.05) | Confirm | Missed bid |
|---|---|---|---|---|---|---|
| eSpeak, noise-free | Deepgram nova-3 | 24 | 1.000 | 0.000 | 0.292 | 0.000 |
| eSpeak, noise-free | Groq whisper-large-v3-turbo with a bid prompt | 24 | 0.923 | 0.000 | 0.333 | 0.000 |
| Human-like synthetic voice, noise-free | Deepgram nova-3 | 24 | 1.000 | 0.000 | 0.292 | 0.000 |
| Human-like synthetic voice, noise-free | Groq whisper-large-v3-turbo with a bid prompt | 24 | 0.923 | 0.000 | 0.333 | 0.000 |
Status PARTIAL. After the evaluation label fix, parser fixes and provider options, both providers are within the limit on both clip sets, but the fixes were chosen on these same clips. The clips are synthetic and noise-free; there is no real-room or noisy-room test. Decision: spoken bids go through the confirm prompt (assist mode). See voice and the confirm gate.
What changes the decision
- G3 shows a validity window shorter than the longest planned auction: auctions are capped, or the decision moves to NO-GO for long auctions.
- G3 passes with a window that covers the planned auctions: GO for card and wallet.
- G5 stays PARTIAL: voice bids stay behind the confirm prompt. The payments decision is unchanged.
- docs/GATES.mdthe record this page is generated from
- docs/evidence/G1raw run logs
- docs/evidence/G2raw run log
- docs/evidence/G4raw run log
- docs/RUNBOOK-GATES.mdhow to re-run a gate
Last updated on