I Funded a Solana Memecoin Sniper Bot. It Went 4 and 82.

Real SOL. An XGBoost model picking every single trade. Fully automated, streamed live. 86 trades, 4 wins, 82 losses. And the most useful thing I learned isn't that it lost — it's why: the bot executed fine. There's just no edge at the moment you have to buy.

This is the whole tape, including the autopsy.

What it is

A Solana memecoin sniper. It watches new pump.fun launches, an XGBoost model (xgb-v3) scores each one in real time on holder/early-window features, and if the score clears the bar it buys ~0.015 SOL (~$1) of the token — then exits on a take-profit, stop-loss, trailing stop, or a ~5-minute timeout. Funded with real SOL out of a real wallet. It runs on camera, 24/7.

The honest scoreboard

This is the "guarded canary" run — the cleanly instrumented window I used to render a verdict:

• 86 trades · 4 wins · 82 losses · 4.7% win rate
• Net: −0.131 SOL (about −$9)
• 74 of 86 trades just timed out — bought a coin, it went nowhere, sold ~5 minutes later for a small loss. 11 hit the stop. 1 trailed out.

And it's worse than "4 wins" sounds, because three of the four wins were rounding noise:

Exactly one trade actually worked — at, which popped fast and trailed out in 38 seconds for +0.0074 SOL (about +50% on the position). The other three "wins" rounded to zero.

Across ~300 live trades lifetime, the wallet went from ~2.26 SOL to ~1.72 SOL — call it −0.48 SOL, ~−$32. Small money, on purpose. The point was never the dollars; it was to find out whether the edge is real. It isn't — at least not where I was looking.

The part everyone gets wrong: it wasn't execution

The obvious story is "you got sandwiched / rekt by MEV / ate slippage." I built a cost-decomposition harness specifically to rule that out, and it did:

• Median buy slippage: +1.21%. Median sell slippage: −1.57%. Fills landed right next to the mark.
• The bot bought when it meant to, near the price it expected, and got out cleanly.

Execution was the best part. (There genuinely was an execution bug earlier — a Solana preflight-commitment issue that failed ~74% of buys outright — and I fixed it. After the fix, fills were clean. The losses kept coming anyway. That's the tell.)

The average win and the average loss were nearly the same size (+0.0022 vs −0.0017 SOL). The problem was never the size of anything. It was the 4.7%.

The actual problem: no edge at entry

Here's the uncomfortable finding. The model rejected 3,037 tokens to select those 86 buys — it was being picky. And the trades it picked still won 4.7% of the time. When I bucketed the trades by the model's own confidence score, a higher score did not mean a better outcome — the score was flat against the result.

Which makes sense if you sit with it. At the instant you're forced to buy — a token 30 seconds old — the information that separates a coin that 10×'s from one that rugs in 16 seconds (hi, METHEREUM: −0.0086 SOL, stopped out in 16 seconds) mostly doesn't exist yet. Holder distribution, dev behavior, real buy pressure — the features that would predict the outcome are still forming after you've already had to commit. The model wasn't broken. I asked it to predict something that isn't yet predictable.

What I'm doing about it

I paused it. The killswitch is on. Throwing $1 at a coinflip with the house against you isn't a strategy.

The questions I haven't ruled out, and won't pretend I have:

• Maybe the edge is in the exit — let the rare winner run instead of timing out at 5 minutes. (at returned ~50% and I clipped it in 38 seconds.)
• Maybe it's in sizing — if the payoff is fat-tailed, the play is more shots, not better entries.

I've trained a v4 model and I'm evaluating whether it actually separates outcomes before I risk another lamport. If it doesn't, this becomes a cleanly documented "no" — and a documented no is also a result.

Why I'm posting the losses

Most trading-bot content shows you the wins. This is the entire tape: 4 and 82, the wallet down half a SOL, the one trade that worked, and the reason the other 82 didn't. I run these on real money, on camera, live, precisely so the failures are as visible as the wins.

The boring grid bot is quietly profitable. The memecoin sniper is paying tuition. Both are true, and I'll keep showing you both.