It's actually trivial to observe with code. Simply add hundreds of irrelevant lines between one relevant variable assignment and its usage, and bugs that were previously detected by the model can no longer be found. With no other changes.

→ View original post on X — @fchollet