Skip to content

JUDGE.md

Latest commit

 

History

History
51 lines (36 loc) · 5.56 KB

JUDGE.md

File metadata and controls

51 lines (36 loc) · 5.56 KB
Raw

For the judge

Live version of this page: https://sentwrong.edycu.dev/judge (no auth, no cookies, no key).

Sent crypto to the wrong address? Paste it. Nansen tells you which of four recovery routes you're on — and drafts the ticket.

The 30-second path — no setup, no key, no clone

  1. Open https://sentwrong.edycu.dev/q/0xe460774c849089ee3edf0fb06da14c066caabbef — a live verdict on a real Binance deposit address: EXCHANGE DEPOSIT · high, four evidence lines, each naming the Nansen endpoint and field it came from, and the prefilled Binance support ticket under Copy.
  2. Open https://sentwrong.edycu.dev/q/0x000000000000000000000000000000000000dEaD — red CONTRACT OR BURN: Nansen refused the address with HTTP 422 "Burn address not allowed", and the card says not to pay anyone who promises recovery.
  3. Open https://sentwrong.edycu.dev/q/0x50b37a3ec6c04609968810801531f5021929eac9 — amber ACTIVE STRANGER · poisoner: an address-poisoning look-alike that only ever "receives" homoglyph tokens.
  4. Go to https://sentwrong.edycu.dev, click the Binance deposit address example and watch the ten Nansen calls land one row at a time; then Every Nansen call (10) opens the provenance drawer: endpoint · fields · credits · ms for every call.

Every verdict above is computed live on the server from Nansen's API when you open the link (the key stays on the server). Expect 4–13 s on a cold Vercel instance — Nansen's per-call latency varies by the minute.

Receipts — real runs, not estimates
Hero verdict, cold, live 13 credits · 10 Nansen calls · 3.9 s (builder, 2026-09-16 15:08 UTC) and 13 s on an independent fresh clone (reviewer, 23:55 UTC); decision hash 3ea6cfcd752b identical on both, and on the recorded fixture
Benchmark (npm run bench) 13 addresses × 3 runs, every call live: 39 verdicts · 309 live calls · 396 credits · 0 non-422 failures; cold p50 3,157 ms / p95 6,477 ms; warm (24 h cache) p50 2 ms, 0 credits, identical hash every run — DEMO.md
Credits spent building it ≈ 2,250 on the shared Meridian key across spike, seeds, bench, QA and screenshots (per-run client sums), plus ≈ 126 in the independent review. The 100-credit labels endpoint was called once, on purpose.
Tests 234 vitest tests in 9 s, of which 24 regression tests named after the defect they pin (e.g. "a transactions timeout read as 'nothing on record'"), plus 13 fixtures recorded live and replayed offline 13/13 with the same decision hash
Property-based verification 40,000 generated Nansen response sets (fast-check, 4 properties × 10,000 runs) through classify() — packages/core/test/classify.property.test.ts: exactly one of the five routes every time; a failed transactions lookup never becomes a stranger verdict; the decision hash is invariant to USD prices; the function is pure. All 14 route/sub-state pairs reached.
Boundary tests The server key never appears in the verdict JSON, the streamed provenance, an error message or the fixtures; every malformed address is rejected with zero network calls — packages/core/test/boundary.test.ts (8 tests) and e2e/key-boundary.spec.ts, run with no key at all.
Timed clean clone install 6 s · live hero verdict 13 s · offline verify 1 s · tests 3 s · dev server first 200 in 4 s (independent reviewer, two clones)

Reproduce — the real path

git clone https://github.com/edycutjong/sentwrong && cd sentwrong && npm install
export NANSEN_API_KEY=nsn_…            # https://app.nansen.ai/api — free tier works, this costs 13 credits
npm run sentwrong -- 0xe460774c849089ee3edf0fb06da14c066caabbef --explain

You get the verdict, the four evidence lines with their Nansen fields, the ticket, and the call-by-call provenance (endpoint · credits · ms · fields read). Web app: npm run dev -w apps/web.

CI / deterministic replay (not the product): npm run verify replays the 13 recorded fixtures with NANSEN_OFFLINE=1 — zero network, zero credits, every decision hash must match. It exists so CI and a reader without a key can watch the engine decide; the CLI and the web app hit Nansen live by default.

Honest limitations

  1. Nansen's per-call latency varies 0.3–3 s by the minute, so a cold verdict takes 4–13 s. On Vercel the 24 h cache is in-memory per function instance — a cold start is fully live. Locally the cache is on disk.
  2. "Nothing on record" cannot tell a brand-new address from one Nansen does not index: vitalik.eth returns empty profiler pages and comes back as a low-confidence fresh stranger. The card says so instead of promising that waiting will help.
  3. Ethereum is the deeply exercised chain; the other seven EVM chains go through the same endpoints but only base has been checked live. The flagship profiler/address/labels endpoint (100 credits) is opt-in (--deep) and Smart Money endpoints are absent on purpose — the beneficiary is a person who lost money, not a trader.

Links