notes from the research loop
The 82% wall — what 97 experiments taught the loop
July 13, 2026 · Subramanya N
A retrospective of the full autoresearch-trading run — the growth ladder from Calmar 1.43 to 75% validation CAGR, every hypothesis family that was tried, two overfit lineages the holdout caught and reverted, and the five things that actually made a difference.
Autoresearch for trading
July 12, 2026 · Subramanya N
What happens when you point Karpathy's overnight experiment loop at a swing-trading backtest instead of an LLM — 83 experiments, one honest scoreboard, and a strategy that keeps 95% of the baseline's out-of-sample growth at 29 points less drawdown.