Chapters on building trading systems that can't lie to their owner, published every two weeks on Substack, with a weekly scoreboard note in between. The failures are where the useful parts live, so they're in.
How to verify AI-generated code when the agent wrote the tests too: independent invariants, planted defects, and a kill count. Measured across five audits.
Chapter oneThe opening move of the whole shop: attack your own results before anyone else can, and publish what the attack finds.
Chapter twoThe bug that made a system look twice as bold as it was, and what an honest post-mortem owes the record afterward.
Chapter threeSurvivorship: most backtests silently delete exactly the companies that would have hurt. Mine kept the corpses. It hurt.
Chapter fourWhat a "years-long" edge turned out to be made of when the window got taken apart day by day.
Chapter sixThe rebuild felt wrong because it was slower. Slower was the point: a walkthrough of the working method.
The graveyard testCorpses and costs left in; the mechanism named honestly. The graveyard wounded it; the tollbooth killed it.
Crash receiptsJane Street lost fifteen billion in a rare bad month. The same weeks, the live book underperformed a monkey. One of us had disclosed the exits in advance.
The weekly scoreboard notes archive on the Notes page, every Monday. The full archive lives on the Substack. Subscribing there is the way to follow along.
The luckiest monkey: what a random baseline did to a beautiful conclusion. The sealed vault opening: out-of-sample day, rules pre-written, answer stands either way. The live experiment's reveal: months of me-vs-monkeys with real NAV. And in December, a long look back at six months of AI agents running a $1M paper book, every mistake on the record.