llm

An AI handed a pile of losing trades will happily invent a lesson from noise. Aegis lets a model propose lessons but never accept them, and a sealed out of sample test does the accepting. So far it has accepted zero, which is the right answer.

You cannot A/B test a trading strategy on one account, because the market is the noise. Aegis runs eight paper accounts that differ by exactly one thing each, so any gap between them is that thing and not luck.

Aegis is a private paper trading harness where the AI is only allowed to veto a trade, never choose one. Here is why I built it that way, and the honest thing it has told me so far.