You already trade the 5-minute markets. You just can’t be awake for all of them.

How much can it lose before something stops it?

None of that is a setting you have to go and find. It is how the file arrives — and you can sit and watch every one of those limits trip in dry-run before a dollar is involved.

Three commands, and it is trading. Nothing of yours is at risk yet.

No wallet. No funds. No account. No API key. Inside a quarter of an hour it is watching real 5-minute windows open and close, taking the trades it would have taken, and writing down what happened to each one — the signal state, 11 market features, the simulated fill. The monitor panel scores it live while you make coffee.

pip install -r requirements.txt
cp .env.example .env
python uruguabot.py        # starts in dry-run, no real money

The .env.example it ships with is not a blank template waiting for you to guess. It is the exact configuration that came out of the live test: BTC only, stop-loss ten cents under your fill, guards on. You can change every line of it. You do not have to change any of it to start.

What it does while you're not watching

Five things run inside that terminal window. Two of them only exist because they cost us money first.

The best predictor we measured was a human

We spent a week settling an argument. Every signal the bot can compute — momentum, mean reversion, orderbook imbalance, the machine-learning model — had its call frozen before each window opened, then scored against what actually happened. 1,934 resolved windows. On the same windows, under the same rule, we also scored the manual picks of the person who built the thing.

Four of the five predictors came back indistinguishable from a coin.

The one that didn’t was the human. 83 picks at 59% gives p ≈ 0.05 against a coin flip — suggestive, not conclusive, one person, one week, and we are not going to inflate it into more than that. But it is the only above-chance directional read anywhere in this project, and it did not come out of the automation.

Which tells you exactly what this is for. It is not sold on out-predicting you. It is sold on the part you cannot do by hand at 4am: sitting on every qualifying window without getting bored, and cutting a loser at −10¢ the moment it hits, without negotiating with itself the way you did that night you held on because it felt due. That discipline is what made the default configuration profitable at a 40% win rate. Whatever read you have stays yours.

So the monitor panel scores your calls next to the bot’s, on live markets, with no wallet attached and nothing at stake. You may find out that your read is the best signal in the room. You may find out it isn’t. Either answer is cheaper to get here than with money on it. Full methodology and data.

The configuration that made money won fewer trades than the one that lost it

That is not a typo, and it is the entire mechanism. In May 2026 Uruguabot traded real money on Polymarket — not a backtest — running two exit configurations back to back. The first was right more often and finished down. The second was right less often and finished up. Computed from the bot’s own fill log, every order published row by row in the audit repository:

• Being wrong more often is what made it work. The profitable configuration won 40% of its trades. Its losers were cut at −$1.09 on average; its winners were held all the way to $1.00 resolution and came in at +$2.31. Roughly two to one. At 40% accuracy that arithmetic clears; at 49% with symmetric exits it did not. The asymmetry is the product.
• The most expensive thing we hit was not the market. Polymarket’s positions API lags on 5-minute windows. Orders that had genuinely filled came back reading zero shares, so the bot wrote them off as ghosts and walked away — leaving real positions open, unmanaged, with no stop-loss on them. Over and over. −$88, quietly, more than the strategy made. The fix — verify the fill against the USDC that left the wallet, fall back to the on-chain balance — ships switched on.
• The machine-learning model never beat a coin. 57% in training, 48–50% once it was live. So it ships untrained and switched off, and you can leave it that way forever. We are not charging you for a number we could not reproduce.

The whole story, every order, every bug, in the order we hit them: the complete audit. And if you would rather read the thing than read about it — 6,735 lines of Python, nothing compiled, nothing obfuscated — the file list and a real function out of the execution layer are sitting in the docs, before you pay.

Before you decide

What $99 gets you

A bot you own, not a subscription you rent

What arrives when you pay

Who should not buy this

Three of them. If any one is a no, close the tab — we would rather lose the sale than take $99 for something that will not run on your machine or in your country.

1. Python 3.11+, on Windows, Linux or macOS. No Docker, no build step, no compiler.
2. Polymarket has to be reachable from your jurisdiction. It restricts trading from several, including the United States — check its current restrictions and your local law before paying. There is no mechanism here to evade them and we will not help you build one. If Polymarket does not serve you, dry-run alone is not worth $99.
3. You are comfortable in a terminal and editing a config file. You do not need to write Python to run it; you do to change the strategy. No GUI installer, and nobody to call.

Refunds

How we found all this out

• We lost $270 building a Polymarket bot — the complete audit

  Ghost fills, phantom A/B gains, orphan adoption, and what finally worked. Every number published.

• We measured a human against 4 trading signals for 1,934 Polymarket windows

  Every automated signal was a coin flip. The human hit 59%. Full data.

• How to automate Polymarket trading in 2026 — an honest technical guide

  CLOB V2, proxy wallets, geographic restrictions, slippage reality, and the bugs that actually cost money.