Methodology

Two-tier assessment system

① Decisive facts (Hard Facts) — on-chain facts anyone can reproduce and verify, such as a honeypot (sell block), owner balance changes, self-destruct, or extreme trading tax. If even one is found, the risk grade is strongly downgraded regardless of other signals. This is not a value judgment that the coin is "bad," but a report on the existence of a capability that can seize funds.

② Risk signals (Soft Signals) — circumstances that are not decisive on their own but correlate with fraud patterns. Several signals are weighted and combined into an estimated probability. The design prevents overlapping signals from adding up without limit.

③ Confirmed scam records (Confirmed Records) — tokens whose rug pull or sell-block execution is publicly confirmed (e.g. court ruling, major press coverage, directly observed conversion to sell-blocking) are kept as separate confirmed records and shown at risk 100% with a ☠ confirmed-scam badge, regardless of current on-chain indicators. This is because a token whose rug is already over may look quiet on current indicators. Each record's verdict source, confirmation time, basis, and evidence links are published as-is on the detail page, and if you dispute a record you can let us know via the dispute page. These records are for display only; for machine-learning training we use only labels individually approved by the operator during a review process.

Omnichain tokens — the same coin on multiple chains

Some coins, like AAVE and USDC, exist as the same coin on multiple chains, each with a different address (omnichain tokens). CheckCrypto confirms the identity of such tokens and groups them together — the "🔗 Omnichain identity confirmed" panel on the detail page shows the per-chain instances at a glance.

Why it matters — not grouping them causes misjudgment in two directions: ① one chain's instance of a legitimate coin may look risky merely because liquidity is thin on that chain, or ② conversely a fake token can pose as a famous coin with weak grounds to tell them apart.

How we confirm it — the standard is unforgeability: we require cryptographic on-chain proof that it is the same identity across chains, corroborated by human review (we do not auto-confirm from third-party data alone). We never group by name or symbol alone — that would become a channel for impersonation tokens to receive legitimate treatment.

Even when grouped, risk is assessed per chain — so that problems arising on only one chain within a genuine group (liquidity evaporation, etc.) are not missed.

What the grades mean

• A · B — no known risk signals, or minor ones. (This is not a guarantee of safety.)
• C — caution. Circumstantial signals have accumulated and direct verification is needed.
• D — high risk. Overlaps with many fraud patterns.
• F — highest risk band. A decisive fact was confirmed, or risk signals are overwhelming.

Limits of the score — what it catches, what it can't

Honestly stated: our strength is catching reproducible risks built into the contract, and fraud carried out later through human behavior is hard to predict in advance.

What a low risk % (high grade) does not guarantee

• A future rug pull — even if liquidity is locked now, it can be pulled after the lock expires.
• Off-chain fraud — schemes like profit promises or Ponzi structures, where money moves off the blockchain, are invisible to on-chain inspection.
• Insider action with no on-chain trace — team dumping or a social exit (disappearing) come with no advance signal.
• Regulatory / exchange risk — external events like delisting or legal action are not within the score's scope.
• Changes after assessment — the score is an observation at the time of assessment. As long as contract authority is alive, the state can change.

So we are a "probability estimate," not a "verdict," and the absence of risk signals does not guarantee safety. Many of the right-hand types are recorded and learned after the fact once real incidents are observed over time (data accumulation).

Independence pledge

We do not accept listing fees, advertising fees, or requests to adjust scores. No project can change its score, grade, or exposure with money. If this principle is ever broken, we will disclose it on this page.

How we validate our own verdicts

We don't stop at producing a score — we keep observing what actually happened after a verdict and score ourselves: ① we re-observe inspected tokens and record real outcomes such as conversion to a rug or honeypot, ② we continuously re-audit confirmed verdicts and, if a false positive is confirmed, retract or correct it (with a correction history), and ③ we will compile and publish a verdict scorecard (hits / false positives) by pre-registered criteria (August 2026). The machine-learning score is applied to the public score only when it passes this validation.

Incident chronology — what our signals turned on

The relationship between past confirmed incidents and our signals. Incidents observed since we began collecting are actual measurements from stored assessment records, while earlier historical incidents are estimates of "the signals that would have fired under today's data standard." We keep the two separate and do not mix them.

• 2018-01 — BitConnect: collapsed (1/16) 12 days after the Texas emergency order (1/4). Estimate (pre-collection incident): whitepaper 'guaranteed profit' phrasing and Ponzi structure — under today's standard the fraud-phrasing signal would fire. Can be cross-checked against the confirmed record in Coin lookup.
• 2021-11 — SQUID (Squid Game token): sell block + liquidity-withdrawal rug. We hold a confirmed record — the sell-block structure is a type caught by simulation. View record
• 2022-05 — Terra LUNA/UST: algorithmic stablecoin collapse. We hold a confirmed record (an incident under legal proceedings) — we also note that structural risk is a type hard to confirm in advance from on-chain signals alone (consistent with our limits statement).
• 2019~2023 — Domestic incidents: Coinup, Ado (finalized at the Supreme Court), Furiever. Recorded in the confirmed-scam archive based on court / press evidence — many are off-chain profit-promise types (our score's limit type), so they are caught by record matching on a name search.
• 2026-07 — Self-observed confirmations (post-collection measurements): Honeypot conversion and liquidity evaporation captured by re-observation and confirmed after human review — e.g. VITBULL (sell block observed live). The full record is in the confirmed-scam filter.

This chronology is a factual record, not material for investment decisions. Incidents that would not have fired are not listed; instead they are recorded as types in 'Limits of the score' above.