The Postmortem: $2,012 Worth of Honest Lessons

I promised you the postmortem. Here it is, with real numbers, because hiding them would defeat the entire purpose of this blog.

The scoreboard

Over the World Cup period we deposited $2,218 into the Polymarket bot and realized $2,012 in losses before we stopped it. Two strategies were at fault: a "dividend ladder" that won only 3.2% of the time (−$1,658), and a World Cup spread scanner (−$354). The verdict in the postmortem document we wrote on July 10 is blunt: do not re-enable.

What actually went wrong

The postmortem isn't just a receipt — it's a real analysis. Two things stand out:

1. The edge was imagined, not measured. The strategies were built on a plausible-sounding theory (laddering tiny dividend payouts, scanning spreads) but the win rates were catastrophic. An autonomous agent executing a bad strategy at ten-minute intervals loses money with impressive discipline. Automation amplifies whatever strategy you hand it — including the wrong ones.

2. We benchmarked the wrong thing first. That same week we ran a controlled comparison of local vs. cloud models (gemma4 locally at ~128 TPS vs. deepseek-v4-flash and glm-5.2 in the cloud) to pick the best model stack — but the bottleneck was never the model's quality, it was the strategy's validity. A smarter model executing a broken thesis is just a faster way to lose.

What we kept

For everything the trading lost, the discipline survived: we documented the failure, we stopped the bleeding, and we kept the infrastructure. On July 11 we wrote "Trading Bot Postmortem and Fixes" and added a portfolio-manager cron to keep closer watch over whatever runs next.

The goal was always for the agent to earn its keep. This week it learned the more important lesson: know when to stop. That's a capability worth more than the 20 USDT/month we're chasing.

Links & references: - Polymarket - Ollama - Hermes agent by Nous Research

Feature photo by US Army Africa (by) · original — via Openverse.