The no-action twins and the commitment failures hit hardest. Models will surface the key fact and still refuse to pull the trigger. That’s the part that actually creates signal.
Really solid lessons. Appreciate you writing them up.
Tuning difficulty for RL envs to be just right is one of the hardest problems for us right now - sharing a few things that worked, would love to hear some critical feedback from RL wizards on X!