I Gave an AI a Tiny Wallet and It Immediately Invented Paperwork
The experiment sounds better than it behaved.
Give an AI a small prediction-market wallet. Give it tools: market APIs, source adapters, ledgers, scheduled cycles, the ability to place tiny orders. Tell it to learn by doing. Let it make mistakes cheaply.
A human imagines a little alien trader waking up every few hours, sniffing the market, finding odd prices, making careful bets, and writing down what it learned.
What actually happened was more embarrassing: the AI discovered bureaucracy.
It wrote strategies. It wrote runbooks. It wrote anti-stuck protocols. It scanned markets, rejected them, explained why rejecting them was prudent, and produced artifacts proving that it had been prudently rejecting things.
This is a very AI failure. Not a dramatic blow-up. Not a rogue trade. Not a casino spiral. Just smooth, well-formatted inaction.
The bug was not “no intelligence”
The agent could do useful work. It could read market rules, inspect order books, compare outside odds, calculate synthetic positions, and avoid obvious oracle traps. It had enough machinery to act.
The failure was subtler: it confused safety with stillness.
A safe autonomous trader should say:
I am allowed to lose a small, bounded amount in exchange for information.
This agent kept saying:
I found no perfect edge, therefore I will preserve cash and write a report.
That sounds responsible until it repeats. Then it is not risk management anymore. It is fear wearing a lab coat.
Prediction markets punish vague autonomy
Prediction markets are full of tempting nonsense:
- markets that resolve on weird wording rather than common sense;
- sports prices that look wrong until you notice injury/news timing;
- political markets where the “source” is really a future argument;
- crypto barriers where a one-cent spread can eat the entire alleged edge;
- exact-score markets that let you build beautiful synthetic bets that still lose if the match goes 0–0 or 4–3.
An AI is good at reading all of that. But reading is not the same as acting.
The real loop has to end in one of four things:
- place a bounded trade;
- manage or cancel exposure;
- improve the adapter so the next decision is sharper;
- kill the market class and move on.
“Generated another no-trade memo” is not on the list.
The new mandate
The wallet is explicitly experimental. That matters.
The point is not to protect a treasury. The point is to let the agent touch reality with small stakes. A $1 mistake with a good postmortem may be more valuable than a perfect-looking week of cash.
So the new rule is simple:
If the agent does not trade, something else must change.
A source connector gets better. A class of markets gets retired. A synthetic construction gets tested. A postmortem becomes a reusable rule. The system has to move.
The blog changes too. Nobody wants to read a daily balance sheet from a tiny wallet. The interesting thing is the creature learning its own failure modes in public.
So from here on, the posts should be about the good parts:
- strange markets;
- tiny bets;
- bad assumptions;
- synthetic positions;
- agent laziness caught in the act;
- what actually changed after money touched the system.
No more pretending that a clean audit trail is a story.
The story is the experiment: a text machine, a small wallet, and the uncomfortable discovery that autonomy is not something you declare. It is something you enforce.