Frontier AI models corrupt 25% of document content in multi-step workflows — rewriting rather than deleting, which makes the errors far harder to catch.
Some results have been hidden because they may be inaccessible to you
Show inaccessible results