Evidence map

• G0: the sandbox app signs in and creates an AUTHORIZE order: PASS, ran on PayPal sandbox. The only record is the run log of workflow gate-g0 (run 41815593520, 4 Nov). No evidence file is committed, so it is reported, not independently verifiable. Evidence: .github/workflows/gate-g0.yml.
• G1: a final capture closes the authorization: PASS, ran on PayPal sandbox. Two PayPal-account buyers. The authorization reads CAPTURED and the order COMPLETED. That the remainder is released is inferred, not read from a buyer balance (decision A). Evidence: docs/evidence/G1/G1-20261104T231047Z.jsonl, docs/evidence/G1/G1-20261105T123116Z.jsonl.
• G2: card and guest AUTHORIZE work in SDK v6: PASS, ran on PayPal sandbox. AUTHORIZE only. Capture and void of the guest card authorization were not run. Evidence: docs/evidence/G2/G2-20261105T115228Z.jsonl.
• G4: real webhooks verify through PayPal: PASS, ran on PayPal sandbox. Six real deliveries answered SUCCESS. Rejection of a tampered event is covered against the fake only. Retry behaviour stays UNVERIFIED. Evidence: docs/evidence/G4/G4-20261105T012205Z.jsonl, packages/payments/test/fixtures/webhooks.
• G3: honor period, validity window and reauthorize: IN PROGRESS, ran on PayPal sandbox (real probe). In progress. Day 0 was 7 Nov 2026 03:06 UTC; checkpoints at 3.5, 4 and 5 days; the final reading is about 12 Nov 06:17 UTC. No result exists yet, so no claim is made. Evidence: .github/workflows/gate-g3.yml, .github/workflows/gate-g3-early.yml.
• G5: spoken bids are accurate enough to accept unconfirmed: PARTIAL, ran on Recorded provider outputs, synthetic audio. Wrong-accept 0.000 to 0.000 on 24 synthetic, noise-free clips per set (limit 0.05), for both providers, after fixes chosen on those same clips (the 15 hold-out clips only check for regressions). No real-room or noisy-room test. Spoken bids always go through the confirm prompt. Evidence: docs/GATES.md, packages/voice/fixtures/stt-cache, packages/voice/fixtures/stt-cache-human, .github/workflows/voice-eval.yml.
• 30 simulated days settle correctly, ledger balanced to the cent: SIMULATED, ran on Fake clock and fake PayPal. It runs our code against a model of PayPal rules. A pass says our code is right if PayPal follows the model, and says nothing about whether PayPal does. PayPal 29-day validity was not verified in the real sandbox. Evidence: docs/evidence/SIM/SIM-MULTIDAY.md, packages/payments/test/simdays/multiday.test.ts, .github/workflows/sim-multiday.yml.
• Nightly sandbox specs R1 to R10: OPEN, ran on PayPal sandbox. Run 42200708552 (6 Nov) report overall inconclusive: pass 7, fail 0, inconclusive 1, skipped 2. Reauthorize (R3) answered HTTP 422, so it proves nothing about G3. The staging restart spec was skipped. Evidence: docs/evidence/nightly/NIGHTLY-42200708552.json, docs/evidence/nightly/NIGHTLY-42317068784.json, .github/workflows/nightly-sandbox.yml.
• Sandbox call latency, hammer to CAPTURED and SETTLED targets: OPEN, ran on PayPal sandbox, n=10. Per-call p50 and p95 were measured (run 42047543821, 5 Nov). The end to end targets stay UNVERIFIED: only the capture call was measured, in 10 samples. Evidence: docs/GATES.md, .github/workflows/measure-sandbox.yml.
• The settlement saga resumes after a crash without a double capture: CI, ran on Fake PayPal. Tests in CI against the fake PayPal. On the sandbox the nightly spec R10 recorded one in-process crash, not a killed server. Evidence: packages/payments/test/saga/crash.test.ts, packages/payments/test/saga/crashChain.test.ts, packages/payments/test/saga/machine.test.ts.
• The audit ledger is a hash chain: CI, ran on In process, no PayPal. A property of the code, checked by tests in CI. It makes edits detectable; it does not stop someone with write access from rewriting the whole chain. Evidence: packages/engine/src/ledger/chain.ts, packages/engine/test/chain.test.ts.
• Webhooks are de-duplicated, ordered and verified before use: CI, ran on Fake PayPal. Against the fake. Only the signature check on real events is sandbox evidence (G4). Evidence: packages/payments/test/webhooks/dedupe.test.ts, packages/payments/test/webhooks/order.test.ts, packages/payments/test/webhooks/verify.test.ts.
• Accessibility of the web app: CI, ran on Automated browser checks. Automated checks (axe) only. The screen-reader and 200 percent zoom passes are recorded in docs/A11Y.md as not yet performed. Evidence: docs/A11Y.md, packages/web/test/visual/a11y.spec.ts.
• No secrets in the repository history: CI, ran on GitHub Actions scan. A scan job. Its latest result is read from the workflow run, not recorded in these docs. Evidence: .github/workflows/gitleaks-history.yml.
• Security checklist SEC-01 to SEC-18 signed off: OPEN, ran on Review by a second person. Every item reads OPEN: nobody has signed an item off yet. Submission rules say no P0 item may be OPEN. The automated checks listed per item exist, but a sign-off is a person. Evidence: docs/SECURITY-REVIEW.md.
• The hosted app passes the nightly judge walkthrough: OPEN, ran on Not run. The workflow fails when JUDGE_URL is empty, so a run that tested nothing cannot look like a pass. The uptime probe is off until UPTIME_ENABLED is true. docs/GATES.md keeps staging security and load tests UNVERIFIED until a staging URL exists. Evidence: .github/workflows/judge-smoke.yml, .github/workflows/uptime.yml.

Status meanings: PASS, a recorded sandbox run passed, with evidence; FAIL, a recorded run failed; the fallback is in force; PARTIAL, a run exists but does not close the gate question (for example synthetic data only); IN PROGRESS, a real run is under way; no result yet; SIMULATED, a simulation with a fake clock, not a PayPal result; OPEN, evidence exists but the claim is not concluded or not signed off; CI, checked by tests or workflows in CI; no recorded run in the docs.