One tweet blew up six of my Polymarket trades overnight. The model was never the problem.

The trade that blew up while I slept

One night this summer, a single Trump tweet closed six of my trades at once. Every one of them an Iran-related bet on Polymarket. My stop-loss did exactly what I told it to and shut each position at -30%, in the dark, while I was asleep.

That one cluster flipped the entire book. With it, forty trades sat at -0.9%. Without it, the same forty sat at +0.85%. One correlated event, one night, and a profitable system turned into a losing one.

The uncomfortable part is that nothing in the machine malfunctioned. The model did its job. I had not done mine.

The machine that made money for zero dollars

Here is what the machine is. It reads Twitter all night with local language models, around the clock: influencers, news outlets, government accounts. A local pass filters the noise down to a signal. A second local pass reads the sentiment, turns it into a probability, and sets that against the live Polymarket price. The automation stops there, at a ranked trade suggestion on a page. I make the final call and place the trade. The human in the loop is by design.

It runs for zero API cost. The pipeline pushes roughly 2 billion tokens a month. On a frontier model that is about $10,000 a month, every month, which kills the economics before you have a strategy. So I moved it off the cloud entirely: a 26B-parameter Gemma 4 model, served locally with oMLX on one Mac with 128GB of unified memory. Not cheaper. Zero.

And under normal conditions, it worked. Across 34 trades it held a 69% win rate and made money. That was the encouraging part, and also the trap.

Why one event took out six trades

The failure was not the classifier. It was correlation.

Six of my open positions were really the same bet wearing six different tickets: the direction of one geopolitical story. A single tweet moved all of them together, and my only defence was a per-position stop-loss firing independently at -30%. Six stops, one cause, one night. Nothing in the system knew that those six lines were one exposure, so nothing capped them as one.

A fast, accurate model bought me good entries. It did nothing to size the book, cap correlated risk, or survive an overnight gap. Speed and accuracy were never the bottleneck.

A fast, accurate classifier without a risk layer just loses money faster.

So I changed the game

The obvious fix is a risk layer that understands correlation and caps a cluster of positions as one exposure. I might still build that. But the blowup pushed me somewhere more basic: if correlated geopolitical bets are the thing that kills you, stop making correlated geopolitical bets.

So I moved the whole system onto up/down binary options on Polymarket. Short, simple markets: will this be up or down at the close of a five-minute window. Two properties matter. Positions no longer all live or die on one headline, so a single event cannot take out six at once. And the trade lifecycle is minutes rather than days, which hands me hundreds of independent outcomes to learn from instead of a handful, and lets me iterate on the system fast.

The real question

Which leaves the question this whole series is about. Can one person run this sustainably and systematically, on that cleaner ground? Not “does the model work.” It does. The open question is whether a fast classifier on short, uncorrelated markets can compound instead of blow up.

I do not know yet. So I am going to find out in the open. 40 days from now, by the end of August, I will have an answer. Real capital, real trades, posted as they happen, the green weeks and the red ones. No cherry-picking, no quiet reruns until the numbers look good.

The easy version of this is to rebuild in private and only show you the winners. I am doing the opposite. Every week I post the number, green or red, until the deadline.

What next week covers

Next, the autopsy. I will walk through the exact cluster of trades that blew up, show why six positions were secretly one bet, and why the overnight stop-loss cascade made it worse rather than better. That failure is the reason the system now trades five-minute up/down windows instead of week-long geopolitical narratives.

If you were betting on me, what breaks first: the signal, the discipline, or the P&L?