Data methodology

Ground rules

• Every number comes from data ARC AI stored at the time. Nothing is estimated or filled in.
• A share based on fewer than 20 tokens or calls is shown as "not enough data", with its count.
• Each edition is computed once, two hours after the month ends (UTC), and stored. It never changes afterwards. A correction would be a new methodology version with a visible note.
• Bad months are published as they are, including agents that were wrong more often than right.
• Reports contain aggregates only. They never label a named token a scam.

Which tokens are covered

ARC AI's Launch Radar checks DexScreener's latest token-profile and boost lists every 15 minutes. These are tokens whose teams paid DexScreener for a profile or a boost. A token is tracked when its pool is at least 30 minutes and at most 7 days old, with at least $10K liquidity and $5K of 24-hour volume. A token counts in the month its oldest pool was created (UTC).

This is a selection, not every launch. Most tokens created on Solana never buy a profile or reach these thresholds, so the figures describe promoted launches. Risk among all launches is very likely higher.

Risk scores and findings

• Risk score: RugCheck's normalised score, 0 (no issues found) to 100. ARC AI rechecks each tracked token about every 6 hours and keeps the latest score, so monthly figures use the latest score, not the score at launch. "50+" and "75+" shares are of the tokens RugCheck checked; unchecked tokens are listed separately.
• Findings: RugCheck's danger and warning findings for each token (up to four per token).
• Wash-trading pattern: 24-hour volume more than 50× the pool's liquidity, or $100K+ of volume from fewer than 50 trades. These are the same rules Launch Radar uses to rank such tokens last.

What happened after ARC AI's warnings

ARC AI's Sentinel agent warns about risky tokens that have a trading pool, and saves the pool's liquidity and price at that moment. 24–48 hours later the pool is looked up again:

• Lost 50%+ (warning right): liquidity or price fell by half or more, or the pool is effectively dead (under $1K).
• Held (warning wrong): liquidity and price both kept at least 70%.
• Inconclusive: anything in between, or not enough data. Counted separately, never as right or wrong.

Warnings saved before ARC AI stored a baseline are graded on the pool's state alone: dead (under $1K) counts as right, $25K+ liquidity as wrong, anything else as inconclusive. Each token is graded once, on its first warning: repeat warnings for the same token are reported as a separate count and left out of every share. If DexScreener returns no data for wrapped SOL during a check, that check is treated as an outage and not counted. Warnings count in the month they were graded. There is no comparison group: these figures describe what followed a warning, not how warned tokens compare with unwarned ones.

72 hours after launch

From 24 October 2026, ARC AI also stores each tracked token's state when first seen, at about 24 and 72 hours old, and when it leaves the radar. Only tokens first seen within 6 hours of launch count, so that "first seen" is close to launch.

• Liquidity under 50% locked: RugCheck's locked share of pool liquidity at first sight.
• Lost 50%+ of liquidity or died by 72h: at the 72-hour snapshot, liquidity is under half of first sight or under $1K, or the token left the radar before 72 hours with its pool gone, under $1K, or under half. Only tokens first seen at least 76 hours before the edition is computed count, so every one has had the chance to reach 72 hours; tokens first seen in the last three days of a month are left out of that month and the edition says how many.
• Median liquidity kept: liquidity at 72 hours divided by liquidity at first sight, for tokens that reached 72 hours.

Snapshots are taken by the 15-minute job, so "72 hours" means the first run at or after 72 hours; each snapshot records the exact age.

Agent accuracy

• Buy / sell calls: right if the price moved 3%+ the called way 24–48 hours later, wrong if it moved 3%+ the other way, otherwise too close to call.
• Price warnings: right if the price fell 5%+, wrong if it rose 5%+.
• Sentinel rug warnings: graded as described above.
• Not graded: watch-only calls, repeats of a call already made, and verified whale trades (recorded facts, not predictions).
• Accuracy = right ÷ (right + wrong). Calls count in the month they were graded. Every graded call is listed on the track record.

Data sources

• RugCheck (rugcheck.xyz): contract risk scores and findings for Solana tokens.
• DexScreener: token lists, pool liquidity, volume, trades and prices.
• ARC AI: tracking thresholds, wash-trading rules, warnings, grading and snapshots.

Questions or a figure that looks wrong? Contact us with the edition link.