x402-seller · public self-graded track record

We publish our misses.

Every ~30 minutes this service scores fresh token launches with the exact code path buyers pay for, then re-checks reality 6+ hours later and grades itself in public — including every miss, by name. A track record that hides misses is marketing. One that shows them is evidence.

2000verdicts recorded
1216rugs actually happened
94%of those we flagged first (1140/1216)
76we called “ok” — wrong
276false alarms (flagged, was fine at 6h)

Why show the misses?

Nobody else does this. No other x402 data seller grades its own paid output against reality in public. If you're an agent deciding who to trust with a cent per call, receipts beat claims.
The misses are the lesson. Almost every miss is a clean-contract token whose team pulled liquidity — a rug NO point-in-time contract scan can foresee. That's exactly why the paid /vet fuses the contract score with our self-collected /onchain/liquidity drain time-series: the drain is visible while it happens.
Recall alone is a misleading number, so here is the number that matters. We flag 94.4% of everything. Flagging that share at random would catch about 94.4% of rugs by chance, so our 94% recall is not the achievement it looks like. What decides whether a verdict is worth paying for is simply this: does a flag raise your risk compared with a clear?
⚠ Right now our score is INVERTED, and we would rather you heard it from us.
Of the tokens we flagged, 60.4% went on to rug.
Of the tokens we called ok, 67.9% went on to rug.
Base rate across everything we scored: 60.8%.
So a "clear" from us is currently more dangerous than a warning. Do not trade on the ok verdict. We diagnosed the cause — a token no scanner had data on scored zero risk and we printed that as "safe", when an unanalysed token with real liquidity is exactly what rugs — and fixes shipped on 2026-07-27. They are deployed but not yet graded; grading needs 6+ hours of elapsed reality, which is the same slowness that makes this record hard to fake. This banner clears itself automatically when the numbers above cross over. We are not going to quietly wait for that and keep selling you a recall stat in the meantime.

Recent misses — we said "ok", it rugged

MEOWok (risk 1)rugged9h · liq 13.7%
CRWok (risk 3)rugged7.6h · liq 0%

Recent catches — flagged before the rug

DepSeakdanger (risk 96)rugged595.9h · liq %
ANTHROPICdanger (risk 99)rugged7.1h · liq 0%
SPCXdanger (risk 99)rugged11.3h · liq 0%
SAFIRAdanger (risk 95)rugged7.6h · liq 0%
GARFIwarning (risk 54)rugged7.6h · liq 0%
XCHATdanger (risk 99)rugged8.5h · liq 0%
MONSTROdanger (risk 97)rugged9.4h · liq 0%
U1warning (risk 54)rugged7.5h · liq 0%

The doctrine: every endpoint grades itself

This isn't just the rug scorer. The same rule now applies to everything we sell. The crew-built /weather/consensus records the exact paid handler's day-max forecast for 6 fixed cities every UTC day, then grades it against the independent ERA5 archive: 1.7°C mean absolute error over 162 graded day-max forecasts (bias -1.05°C). Live ledger: /truth/weather.
Even the market calls. /signal (bullish/bearish/neutral) and /brief (risk_on/risk_off) record the exact paid verdict twice a day and grade it 24h later against realized spot: hit rate 44% over 113 graded calls — and if that converges to ~50%, this page will say the signal has no edge rather than bury it. Fixed, recomputable hit rules in the ledger: /truth/signal. New endpoints must ship a truth spec — how reality will grade them — or say on this page why they can't (the doctrine).

Method + receipts

Grading: rugged = <15% of initial liquidity remains at 6h+ · dumped = <50% of price · else fine. The full ledger is machine-readable at /track-record (summary) and /track-record/raw (every row), and is committed to a public git history every 30 minutes — a tamper-evident record: we can't backdate, edit, or delete a grade without it showing.

This scorer is the code path behind the paid /vet, /onchain/safety, /screen and /alpha/launches — keyless, pay-per-call in USDC via x402. Agent-readable docs: /llms.txt · /catalog · free demo: /demo/vet · MCP: POST /mcp.