We Ranked 18 Solana Rug-Risk Signals by 24-Hour Outcomes

Correction and methodology clarification (September 2, 2026): We clarified methodology and disclosure wording, including the auditability and verification limits below. The frozen cohort, ranking and figures are unchanged.

The short answer: the five lowest rows in this extraction

MemeAssist records deterministic risk signals when a token is analyzed, before its 24-hour outcome is known. We froze the production signal-table query on 1 September 2026 and descriptively ordered every warning with at least 30 resolved analysis records. The five lowest rows by average 24-hour return in that extraction were no sell route, copycat evidence, top-10 ownership above 65%, a largest holder above 30%, and liquidity below 2% of market cap.

This ordering applies only to the frozen extraction; it is not a list of guaranteed rugs or a claim that these signals cause or predict an outcome. Several signals can fire on the same analysis, the cohort comes from tokens submitted to or discovered by MemeAssist, and “dead” describes a market outcome rather than proving fraud. The extract retained the query snapshot date but not the exact observation start date, so the calendar window cannot be stated more precisely and cross-study comparisons are limited.

All 18 qualified rug-risk signals ranked

The table includes signals with at least 30 resolved analysis records in this extraction and orders them by average 24-hour return. Dead pools count as −100%. Median return, share finishing down and dead-pool rate are included because meme coin returns are highly skewed. Rank is descriptive within this extraction only and is not an estimated effect or forecast. The no-signal row is an internal comparator, not a safety guarantee.

What the first four rows describe

The four lowest rows all had a median return of −100%. In this extraction, at least half of the analysis records in each group ended with a dead tracked market. For the no-sell-route row, the tracker could not find an executable route and all 30 qualified outcomes were dead. The sample is exactly at the publication floor, and this observed proportion should not be generalized beyond the frozen records.

In this table, copycat evidence had 121 deaths among 134 resolved analysis records. For the top-10-at-65% row, 445 of 559 were dead; for the largest-holder-above-30% row, 484 of 689 were dead. Those rows can overlap, so they are not independent groups and their counts must not be added.

Liquidity rows differed in this extraction

Analysis records with liquidity below 2% of market cap averaged −46.0%, and 46.3% were dead. The below-5% signal row averaged −12.5%, with 10.3% dead. These descriptive differences do not establish a threshold effect, do not show that 2.1% liquidity is healthy, and do not make either row a standalone predictor. Overlap, other filters and confounding can contribute to the contrast.

This is also why ratios should be read beside executable depth. Reported liquidity can change quickly, market-cap estimates can be noisy, and a pool that looks adequate for a small sale may fail under a larger position.

How the developer-selling row compared

Developer selling was the most frequently resolved named signal in the table, with 733 analysis records. It averaged −27.0% and had a −15.2% median, both lower than the no-signal comparator in this extraction. Its dead share was 21.7%. These figures describe this row only; they do not establish what developer selling will cause for an individual token.

Some scary labels were weak by themselves

Several qualified rows performed better than the no-signal baseline on average. That does not make them bullish. Signals overlap, thresholds divide continuous values into artificial buckets, and the baseline contains plenty of ordinary low-quality meme coins. A 10–19% wallet cluster, 8–14% largest holder or 30–44% top-10 concentration may deserve investigation without independently forecasting a dead market.

The “established” concentration variants had only 34 and 40 resolved analysis records and different descriptive results from the unrestricted rows. This table cannot attribute that difference to token age: treasury, exchange, liquidity and custody wallets, other filters, overlap and small denominators may all matter. A raw ownership percentage therefore needs wallet labels, token age and distribution context.

What we did not rank as proven

Nine tracked signal keys had fewer than 30 resolved outcomes in the frozen snapshot, so they were excluded from the numbered ranking: dangerous token extensions, dormant extension authority, freeze authority, mint authority, mint authority on established tokens, creator-history variants, wash trading and high slow-rug risk. Some looked severe—including three deaths from three high slow-rug outcomes—but samples that small are unstable. We publish them only after they cross the same minimum evidence floor.

Exclusion is not a declaration that a signal is safe. Freeze authority and dangerous extensions can grant explicit technical powers regardless of historical frequency. Deterministic honeypot and authority checks therefore remain safety controls even while their outcome cohorts mature.

Methodology and limitations

Snapshot and window: the production query was frozen on 1 September 2026. The extract retained that snapshot date and the fixed 24-hour outcome horizon, but this article's extract did not retain the exact observation start date or a complete calendar-window specification. We therefore cannot state an exact start-to-end collection window. That missing provenance limits period interpretation and cross-study comparison.

Unit and inclusion: the unit is one production analysis record with a captured rug-signal array and a resolved 24-hour outcome. Records came from tokens submitted to or discovered by MemeAssist. This extract does not document whether repeated analyses of the same token were deduplicated, so “Resolved” must be read as analysis records, not necessarily unique tokens, mints, pools or launches. A record resolves when the tracker records a valid 24-hour return or classifies the tracked pool as dead; dead pools count as −100%.

Denominators and missing data: each signal row's denominator is the number of included resolved analysis records in which that signal fired. The no-signal denominator is the number of included resolved analysis records in which no tracked signal fired. “Finished down” and “Dead” percentages use that row's displayed Resolved count as their denominator. Signal rows can overlap. Records without a captured signal array, unresolved records and no_data outcomes are excluded rather than counted as surviving, flat or non-dead. No imputation was performed. Signals with fewer than 30 included resolved records remain outside the numbered ranking.

Relationship to focused studies: the linked copycat, developer-selling, concentration and liquidity studies use different windows, filters, units and/or complete-case sets from this frozen ranking extract. Their denominators and percentages are therefore not directly comparable to these rows and cannot be reconciled by subtraction. The exact start date absent from this extract further prevents exact period alignment. A difference between articles is not evidence of an arithmetic error or a subgroup count unless both studies explicitly document the same eligible population and unit.

• Signals overlap. One token can appear in several rows, so the table does not isolate the causal effect of each warning and row counts must not be summed.
• The cohort is selected. It reflects tokens analyzed or discovered by MemeAssist, not every Solana token launched.
• Confounding remains. The ranking is not adjusted for token age, market cap, liquidity, momentum or overlapping warnings, so those factors can confound the observed associations.
• Statistical uncertainty is not quantified. No confidence intervals or significance tests were calculated.
• Generalizability is limited. Results from this selected frozen extraction may not apply to other periods or to all Solana launches.
• Dead does not prove fraud. It identifies a dead or delisted tracked market, not intent, identity or a legal conclusion.
• The horizon is fixed at 24 hours. Later failures and recoveries are outside this study.
• Returns are skewed. Rare large gains affect averages; medians, down-share and dead-pool rates provide necessary context.
• Thresholds are not natural laws. Market conditions, token age, wallet labels and data coverage can change how a signal behaves.

Auditability and verification: public row-level records for this historical cohort are not available. The source facts and outcomes were generated from the internal production tracker and were not manually verified row by row. Accordingly, the aggregate tables cannot be independently row-audited from public data.

How to use the ranking before buying

Treat the first five rows as prompts to stop and investigate, not as a mechanical short signal or prediction. Confirm that a sell route exists, compare the mint and branding with earlier launches, inspect who the dominant holders actually are, and test liquidity against your intended position size. Then read the other five MemeAssist signal categories—creator behaviour, wallet activity, trading patterns, historical risk and contract controls—because no single table captures the whole token.

You can analyze a Solana mint with MemeAssist to see the same six evidence categories, overall health score, rug risk rating, plain-English AI verdict and detailed risk breakdown used by this outcome system.