01 / The clearest example
A lab that put a trading edge on trial
A Polymarket wallet was posted online as proof that a bot had cracked five-minute Bitcoin markets, a roughly $21k winner. I did not take that claim, or my own hope that it was real, on faith. I built a read-only lab to reconstruct the wallet from public on-chain data and test whether the edge survived contact with fees and network latency.
- $21k → $1,126 The posted claim, versus the reconstruction after fees
- 52.4% vs 49.8% Actual win rate against the break-even it needed
Reconstructed from the on-chain trades, the wallet made $1,695.71 gross and $1,125.68 after fees, which ate about 34 percent, across 127 resolved markets. A real but razor-thin edge, and the slider above shows what happens to it the moment a real order has to travel.
The verdict: the edge is fee-eaten microstructure, not prediction. Polymarket's fee schedule is built to tax exactly this kind of trade, so the profit belongs to co-located bots, not to the strategy.
Every number here is reproducible. The lab is public and MIT-licensed, with the data committed to the repo. Clone it and the reconstruction runs.
02 / How the work gets made
I direct AI agents, and I make them prove it
Most of what I build, I build by directing AI agents rather than typing the code myself. That only produces something you can trust if the process has structure, and I built mine out of plain text files and git rather than any agent framework.
- A durable backlog the agent reads first. Without it, each session re-derives a plausible, wrong priority and half-finishes the wrong thing. The backlog also records what was tried and deliberately rejected, so no session re-litigates a dead idea.
- A task carries a "done when" line. It names the concrete check: the exact count, the cases that must not fire, the order things deploy in. An agent cannot pass that by feeling good about its work.
- The review gate is a branch, not a permission prompt. Agents work freely on a branch and I review one coherent diff. Nothing reaches the main line unread.
- Capability limits are written into the rules. Where an agent structurally cannot verify something, judging audio, feeling latency on a real phone, that limit is policy, not something it is allowed to fake.
- Risk is enforced where a tool is called, not in the UI. How autonomous an action is and how reversible it is are treated as two separate axes, and the dangerous quadrant is gated.
Here is the one time it was tested for real. I let a batch run unattended overnight. It hit a blocker that would have corrupted live leaderboard data, and instead of pushing through, it stopped and quarantined the unfinished half on a branch for me to review. That is one night, not a fleet of nightly robots. It is the honest version, and the controls held the once it counted.
03 / Other things I've built
A game, a hiring tool, a scam checker
- Wastrel A daily-seeded strategy game with a deterministic integer core. It taught me that determinism is what makes fairness measurable: the same seed and inputs always produce the same result, so anti-cheat and balance become arithmetic instead of opinion.
- Fit or Pass A job-fit tool where a second model audits the first, so it cannot just flatter you into applying. It taught me that on an AI product the evaluator is the product, and the generator is the easy half.
- Is This A Scam A scam and claim checker in ten languages. It taught me that the hard part of a trust tool is saying "there is not enough evidence" instead of guessing, because a confident wrong answer is worse than an honest "I don't know."
04 / Contact
Say hello
Reach me at hi@hugo-marinho.com.