I caught the same wrong fact in three different reports last week.

Numbers were correct. Charts were clean. Conclusions were garbage.

The date range included bot traffic. Nobody flagged it. The knowledge that those weeks were polluted? Lived in my head.

My AI caught it. Because I taught it that fact three weeks ago in one line of text.

FFS, we're spending billions optimizing retrieval. Bigger context windows. Better RAG. Faster inference.

Nobody is asking whether the fact being retrieved is actually true.

Retrieval correctness is not factual correctness.

The fix isn't a better model. It's a text file where you write down what you already know. 30 seconds. One correction. Every future output gets better.

I wrote about how I actually do this, why I think about reports like code commits, and the Sisyphus problem of building context.