prompt phrasing sensitivity

reinforcement learning training traces

both indicate that model behavior is governed by signals outside task semantics — training procedure and input tone — which evaluation harnesses must control

1 source videos