Correction: the September 9 Solana venue comparison is withdrawn

Correction — September 10, 2026

The venue rates, intervals and 7.7x comparison below are withdrawn as population findings. An audit found observations from different sampling designs, including missing inclusion probabilities, capacity truncation and incomplete follow-ups. The newer collector does not retroactively validate those observations.

The original table is retained as a historical record of the calculation, not a current estimate. A pooled rate is also a composition statistic, but splitting by venue does not fix an invalid sample. Missingness and discovery gaps can bias results in either direction; the earlier claim that the true rates must be higher was unsupported.

New reports require an explicit sampler epoch, compatible inclusion probabilities, mature collection sweeps and adequate follow-up coverage. Uncertainty is evaluated across whole sweeps. Until those checks pass, rates and comparative headlines are withheld.

Original text — retained for the correction record

What this was

This is a prospective observation study, which is the opposite of the reconstructions we usually publish. Nothing here was rebuilt after the fact. A collector watched Solana's new-pool feed, recorded every launch it saw as it appeared, drew a random sample of them, and came back an hour later to see what had happened. The numbers below were not knowable at the start of the day.

It is not a rug-pull census, and the word "rug" does not appear in the measurement. What we measure is narrower and checkable: *liquidity below one tenth of what the pool held when we first saw it, one hour after we first saw it*. That captures rugs. It also captures failed launches, migrations, and tokens nobody wanted. Separating those requires intent, and intent is not visible in a reserve balance.

It is also one day. Every figure is a snapshot of 9.4 hours on 9 September 2026, and memecoin venues change quickly.

The finding

Across 347 launches followed to a clean one-hour reading:

pump-fun and pumpswap do not overlap. They are not the same population being measured with noise; they are different populations. A launch on pumpswap was 7.7 times as likely to lose its liquidity inside an hour as one on pump-fun, on the same day, in the same market.

Why the pooled number is the misleading one

17.3% is arithmetically correct and close to useless. It is a weighted average of 7.8% and 60.0%, and the weights are the venue mix — which moves on its own. A day where pumpswap's share of launches doubles produces a higher "Solana rug rate" with no change whatsoever in the risk of launching anywhere.

Anyone quoting a single rate for a chain is quoting a composition statistic and calling it a risk statistic. The two move independently, and only one of them is about danger.

What we did

The denominator is a census, not a sample. Every launch the feed showed us was recorded — 7,913 of them over 9.4 hours, about 842 an hour. Recording a launch costs nothing beyond the page call already made, so there is no reason to ration it, and a rate needs a denominator that was not itself selected.

The numerator is a random sample of that census. Following a pool to an outcome is expensive, so only a fraction can be followed. Each launch faces the same fixed probability, drawn independently — which makes the followed set self-weighting: a plain share over it estimates the population share with no correction step, and the confidence intervals above are computed on the design that actually produced the data.

That sounds like a technicality. It is the whole difference between a measurement and a number. A sample that takes a *fixed count* per sweep gives each launch an inclusion probability of count-over-however-many-launched-that -minute, so launches arriving in a quiet minute are massively over-represented. We measured that on our own earlier data: inclusion ranged from 1.3% to 55%, a 42-fold spread, on a design that looked random at every individual step.

The intervals are Wilson score intervals, which stay inside [0, 1] at small n where the textbook interval does not. A negative share of anything teaches a reader to distrust the rest of the page.

Five venues are absent from the table — meteora-dbc, fluxbeam, raydium-launchlab, raydium and meteora. Between them they contributed 13 followed launches. They are counted in the census and deliberately not rated: below thirty observations we publish the count and refuse the percentage. meteora-dbc came in at five of seven, which would read as 71% and means almost nothing.

What is wrong with these numbers

Stating this plainly, because a study whose limitations are in the footnotes is a marketing document.

Correction to the original limitations. The recorded median discovery lag was 379 seconds. We cannot infer the direction or size of bias from that lag. The old intervals also treated correlated observations as independent; they should not be used as uncertainty estimates for the wider launch population.

One hour is one horizon. A pool that still holds its liquidity at sixty minutes has not survived; it has survived sixty minutes.

The followed sample skews young within each venue. An earlier version of the collector spent its follow-up budget on the first page of the feed, so the sample is drawn from the newest launches of each sweep rather than uniformly. We checked whether that distorted the venue comparison: venue composition on page one versus deeper pages is 65.1% against 67.6% for pump-fun, with every venue inside 2.5 points. Page position is essentially independent of venue, so the venue-conditional rates survive it. The bias is real and it is not this bias.

Why it matters if you are pricing risk

The venue a token launches on is visible at block zero. It costs nothing to observe, it requires no model, and on this day it separated an 8% outcome from a 60% one.

Most memecoin risk scoring we have seen reaches for holder concentration, contract flags and social signal — all of which are downstream, several of which are trivially spoofed, and none of which are free. The strongest observable in our data was the one printed on the pool at creation.

That is not a complete risk model and we are not selling it as one. It is a prior, it is a strong one, and any model that does not condition on it is leaving the most reliable feature on the table.

Reproducing this

The collector is ghostcopy/solana_launches.py and the analysis is ghostcopy/solana_intel.py. Every figure above comes from one command:

python -m ghostcopy.solana_intel

It prints the coverage, the sampling health, the data quality and the per-venue table, and it reports its own problems as findings rather than leaving them for a reader to notice. If the sampler drifts, that output says so before any rate is quoted.

We are continuing to collect. The horizons extend to six and twenty-four hours, and the one-hour figure is the first that has enough observations to publish. When the longer ones do, we will publish those — including if they disagree with this.

If you need this measured on a specific venue, chain, or set of counterparties, that is what a Forensic Sprint is.