Reading risk scores
Every analysis produces a single risk_score from 0–100. It's a weighted, normalized blend of independent risk dimensions — each scored 0–100 on its own, then combined. Higher means riskier.
Risk levels
The numeric score maps to a risk_level and a recommendation:
| Score | Level | Typical recommendation |
|---|---|---|
| 0–25 | LOW | SAFE — appears legitimate |
| 26–50 | MEDIUM | CAUTION — investigate further |
| 51–75 | HIGH | AVOID — multiple red flags |
| 76–100 | CRITICAL | AVOID — strong scam indicators |
The dimensions
The engine evaluates up to twelve independent dimensions, each scored 0–100. Their relative influence is tuned and proprietary; the categories are:
| Dimension | What it detects |
|---|---|
| Owner privileges | Mint/freeze authority, pause, blacklist, balance-modifying admin functions. |
| Holder concentration | How tightly supply is held across the largest wallets. |
| Liquidity | Pool depth, lock status, lock duration, pairing. |
| Ownership transfers | Suspicious owner changes, transfers to fresh wallets. |
| Contract verification | Verified source, audit status, honeypot patterns. |
| Trading patterns | Volume/price anomalies vs. historical behavior, wash trading. |
| Pump.fun signals | Launchpad-specific risk signals (Solana). |
| Supply | Supply structure and inflation potential. |
| Relaunch risk | Copycats / repeated-deployer patterns. |
| Proxy risk | Upgradeable proxies and their centralization risk. |
| Gas anomalies | Unusual transfer-gas behavior (a honeypot tell). |
| Scam connections | Links to known scam addresses. |
How scores combine
Each dimension yields a 0–100 sub-score. The overall score is a weighted, normalized blend of the dimensions that could be evaluated for a given token — so a missing data source doesn't unfairly skew the result. The exact weighting is proprietary and tuned against real-world outcomes.
Confidence & coverage
Responses include a confidence value (0.0–1.0). When too few core dimensions can be evaluated, confidence drops and the recommendation leans toward CAUTION. Treat low-confidence results as "insufficient data," not "safe."