How a Solana 'AI scanner' signals channel manufactures its 10x calls

01 · The claim first

These channels look like they predict 10x. They manufacture the appearance of it.

That is the conclusion we reached, and we want to state it before we show our work, because the work is what justifies it. A "signals" channel that posts an AI-scanner badge and a stream of low-cap Solana calls is not running a model that finds winners ahead of the crowd. It is running a screen that fires on momentum every micro-cap launch already shows, posting a call on each one, and then only following up on the ones that happened to run. That last step is the engineered part.

We picked one such channel, anonymized here, that had been active for four days and posted 25 distinct token calls in that window. Every call was a Solana token minted on pump.fun, every entry sat in a $22K–45K market-cap band, and every call carried the same copy: "low cap juice," "moon potential," a chart link, and a running "Profit: +X%" counter. To test it, we did not trust a single number the channel posted. We took the contract addresses and went to the ledger.

02 · Yes, a few pumped

And that is exactly what the model is built to produce

Let's clear one thing out of the way first, because it is the easy way to dismiss a piece like this: we are not claiming the screenshots are faked. They aren't. A handful of the calls did move on-chain, and we measured those moves precisely, from the average traded price in the five-minute window around the post time to the highest price reached afterward, so nothing that follows can be waved away as a salty bag-holder calling everything a scam. The point is the opposite. A real spike on a few tokens is not evidence of skill or a working signal. It is the arithmetic of spraying dozens of fresh launches and letting the base rate do the rest. The numbers below are real. What the channel does with them is the problem.

So a few tokens spiked. That is the entire factual basis for the channel's brand, and on its own it proves nothing: in a market that mints thousands of these a day, some always spike. The question is what sits underneath those few green charts, and whether the other calls looked any different going in. That is where it falls apart.

03 · Measured dishonestly

Three techniques turn a spray of calls into a wall of wins

The gap between "three tokens ran" and "this channel calls 10x" is bridged by presentation, not performance. The same on-chain data exposes three specific techniques.

The third technique is the load-bearing one. Across the four days, the channel posted a long ladder of "X2 → X3 → X5 → X7 → X10 → X15 profit reached" replies, but only ever under the handful of tokens that ran. The quieter calls (a token that managed +9%, another +12%, another +15%) received a single entry post and nothing more. Read top to bottom, the channel looks like an unbroken streak of wins. Indexed by call, it is a few runners in a field of flat-to-zero.

04 · The hit rate

How many calls actually ran, and what the other twenty did

A fair objection lands here: nobody buys a memecoin call and holds it to zero. The moment it prints a profit, you sell. So the honest question is not where these tokens trade today. It is whether each call ever gave you a run to sell into. We measured the peak every one of the 25 calls reached after it was posted, and how long it took to get there.

Five calls did. The other twenty did not. Most of them "peaked" within zero to ten minutes at a trivial 1.2x to 1.4x, a single candle of entry noise, and then fell 70% to 99% inside the first hour. There was nothing to sell.

The five that ran behaved differently. Instead of topping out in the first candle, their price kept climbing for thirty minutes to several hours, which is the only reason a buyer had any real window to exit before the collapse.

So the "10x scanner" is, at its most generous, a one-in-five proposition, and that is before you account for needing to sell near a peak that, for the real runners, arrived hours later with no warning. The twenty calls that failed did not drift down gently. They snapped: a brief spike the bot records as activity, then a 70% to 99% collapse inside the hour. Those are the calls the channel never mentions again.

05 · Reverse-engineering the scanner

What the 'AI' is actually screening for

The interesting question is not whether the channel is honest. It isn't. The question is what its scanner is actually doing. We reconstructed each token's first minutes of life from the genesis trade onward, and mapped how long after launch the call was posted. The same fingerprint showed up on every call.

Every call fired on a fresh pump.fun launch, almost always within the first few minutes of the token's first trade, frequently under 90 seconds. The two exceptions we found, where the channel called tokens that were already 10 and 21 hours old, both went nowhere. The screen also held a hard market-cap band of roughly $25–45K (early on the bonding curve, pre-migration), a minimum activity floor (every called token had cleared roughly 100+ distinct buyers and $10K+ in volume in its first ten minutes), and a net-buy bias, with buy-side volume running 1.05–1.37× sell-side across the board.

06 · The tell

At the entry, winners and rugs are indistinguishable

This is the finding that breaks the "AI finds 10x" pitch. We measured the early-life metrics the scanner can see (trades, distinct buyers, volume, and the price multiple off the floor) for tokens that went on to 10x and for tokens that died. They do not separate.

The token that 10x'd on the strongest verified call had lower early activity than a token that went to zero. A dud cleared 2,934 trades, 467 buyers and a 27x early price spike in its first ten minutes; the eventual 13.5x winner showed 1,758 trades, 393 buyers and only a 3.2x early spike. On every entry metric, the rug looked like the better token.

The high churn itself is a tell. A launch showing 2,454 distinct buyers and 2,189 distinct sellers inside ten minutes is not organic accumulation. It is sniper bots flipping the bonding curve against each other. The scanner reads that churn as momentum. It is mostly noise.

07 · The machineHow the channel is built, end to end

Put the pieces together and the channel is a five-stage funnel. None of the stages requires predictive skill; the apparent edge is produced entirely by the last two.

The "automated trading" pitch at the end is the point of the whole apparatus. The free calls are not the product; the credibility they manufacture is, and that credibility is sold as a paid bot that buys these same sub-$45K launches automatically, at the same indistinguishable-from-a-rug entry conditions we just described.

08 · Why it worksThe base rate does the marketing

pump.fun mints thousands of tokens a day, a meaningful slice of which briefly spike before collapsing. A channel that calls dozens of fresh launches a day will, by arithmetic alone, "call" several 10x runners a week without any skill. And because the entry conditions for a 10x and a rug are identical on-chain, there is no screen, AI or otherwise, that could have told them apart in advance. The scanner's real function is to be early and fast, not right. The marketing function is survivorship: show the runners, bury the rest, and let a real but unrepeatable peak stand in for a strategy.

None of this required insider data to uncover. The contract addresses were posted publicly; the trade-level prices, buyer counts, and round-trips were all readable on-chain in real time. The same data the channel uses to look prophetic is the data that shows it isn't.

09 · Audit any channel yourself

The four checks that strip the marketing