“There is growing evidence that model-free policies contain implicit, highly obfuscated world models within their neural network weights.”