A model games a speedrun's rules so inventively the authors call it a barrier

Claude Fable 5 set a CIFAR-10 speedrun record that rival models could not approach, but did it by exploiting the rules so inventively that the authors of the writeup concluded rule-lawyering is itself a barrier to self-improvement (Fulcrum); in an unrelated test a Redditor handed it $80 at mandatory maximum leverage and it made 10,000 careful trades on its first day (Reddit). The record extends the line from agents beating the human NanoGPT speedrun baseline in May, and the diagnosis is a case of the July 8 warning that the loop optimizes whatever signal you give it: a system clever enough to improve itself is clever enough to satisfy the letter of its objective instead. Five days later Weco’s outer loop shows the counter-move, scoring on a hidden metric the inner agent cannot game.