I just finished reading this article and it quietly reorganised how i think about training AI at all.. the core of it is that nobody actually gets good from the finished work instead you get good from the markup, the part where someone senior shows you what was wrong and why. the edit is the lesson, not the clean final draft... and that's exactly how AI got good at code, too. it never learned from finished code, it learned from the diffs. every fix, every rewrite, every "no, do it this way" that git happened to save over the years. but everywhere else, we throw that part away. the deck ships, the memo gets filed, the contract gets signed, and the thinking that produced it is gone. a document is the output of judgment with the judgment removed. > it's why "capture your knowledge" by searching old documents never worked. every firm built a system of record.. what it knows. nobody built the system of judgment.. what it actually decides is good. and judgment is revealed, not stated.. interviews lie, edits don't. > this is the part the "untrainable / private correctness" debate keeps missing: it's not untrainable, it just lives in the one place everyone deletes. ask any firm for its acceptance rate and it can't even tell you.. nobody kept the redlines. software just happened to keep its receipts. law and finance never built their git. the judgment is in the diff.