Beta
← Back

Methodology

Data sources

RugCheck (risk flags, holder concentration), Birdeye (holder count, price), Dexscreener (liquidity, volume), GMGN (buy/sell pressure, tax/honeypot), Helius (on-chain deployer history, wallet funding traces for bundling detection). A fresh lookup calls all five; results are cached 15 minutes per coin.

Snapshot hygiene

A 0-100 score built from a fixed set of deterministic rules — no AI involved. Starts at 50, then adjusts for: RugCheck risk score, LP lock status, mint authority, liquidity depth, top-10 holder concentration, holder count, honeypot/tax status, token age, graduation status, and (if run) bundling classification and deployer wallet history. Purely mechanical checks are capped at 85 — only bundling analysis and deployer history can push a score above that, so the top tier means more than "no red flags on the basics."

Community votes and manual review never affect this number.

Outcome prior

MobScore logs a structural fingerprint (holder-count bucket, liquidity bucket, buy/sell-pressure bucket) for every AI-analyzed coin, then checks back 3+ days later to see what actually happened. A coin counts as a confirmed rug if its liquidity has dropped 90%+ from its value at verdict time. The outcome-prior module on a coin's page shows how similar past coins turned out — but only once at least 10 matching past cases have a known outcome. Below that, it says so instead of showing a percentage.

The buy/sell-pressure dimension is only used when that data is available at match time. On a coin's own page it currently never is — the raw buy/sell counts aren't stored on the coin row — so matching there happens on holder-count and liquidity buckets alone. The "Matched on" line in a coin's Similar Past Cases section names the dimensions actually used.

Confidence

Starts at High, drops one level for any core data field missing (RugCheck score, LP lock, mint authority, liquidity, holder count, or honeypot status), and drops another level for an unresolved negative contradiction between sources — each is a real signal about this specific coin. Floors at Low.

Two further gaps — bundle/sniper clustering not having been checked yet, and too few similar past cases yet to compute a reliable outcome prior — are about MobScore's own coverage, not about the coin itself. Bundling now runs automatically on most fresh lookups (see below), so this gap is rarer than it used to be, but it can still happen — a daily budget cap, or a coin with no holder data yet. The second gap is true for nearly every coin right now regardless of quality, simply because the tracked history is still small. Together they cost only ONE combined level, not one each: either or both being true drops confidence by one step, so two purely-informational gaps can no longer out-weigh two genuine per-coin risk signals the way they used to.

Three separate factors are hard caps instead of one-level drops: a confirmed hard bundle, an unknown deployer identity, or a token under 3 days old each force Low regardless of how clean everything else looks. Each removes or fails an entire risk dimension this tool can check, not just one data point — that can't read as anything but uncertain. An unchecked bundling analysis is treated differently from a confirmed hard bundle: not having looked yet is a smaller gap than having looked and found a real problem.

Known failure modes

What the agent can and cannot do

The Mob Agent finds contradictions between sources and flags checklist items that are contextually misleading for a specific token (most often an automatic default on a launchpad token being misread as a real decision). It cannot invent liquidity numbers, override a hard on-chain fail, or endorse a coin.

Changelog