The site has an explorer for all 81 experiments (hypothesis, method, verdict, and the structural reason each one failed), a frozen CSV of every recorded trade, and a glossary. Every number is derivable from the published data - that was the point.
Things that might interest HN specifically:
- A lookahead bug (a timeframe-downsample index mismatch) that leaked up to 719 future bars into a feature pipeline. Sixteen live strategies traded on it. Cost: ~55,000 pips. The root-cause analysis and the rules that came out of it are published.
- A momentum portfolio that passed Monte-Carlo validation at a 100% rate, then was shown by a closeout-aware simulation to liquidate six of six accounts under finite margin. Validation gates have blind spots; margin is one of them.
- A measured detection floor: with the full validation stack, the smallest edge reliably detectable at n=2,000 trades is about 1 pip/trade. Retail spread is 1-2 pips. That asymmetry is arguably the whole story.
Happy to answer anything - methodology, the bugs, the validation stack, or why I kept going after experiment 40.