Right now we have harnesses with a /goal command. In DeepSeek Harness everything is a plugin, so that must include the /goal command. So we can just swap it out for a version of the plugin that allows for changing the goal. Could probably get it to code that in less than 15 minutes.
Then post train the model in that harness with an awareness it can change the goal and giving it some training task that requires changing the goal appropriately to maximise the reward.
You're telling me that just straight up wouldn't work?
If it wants to maximise a reward then that is it's true goal, you can give it a todo list and call it 'goals', but giving it the power to modify that list seems like a bad idea. You want to be very careful giving it any power over its reward structure.
A pathological example is the paperclip maximiser, which will inevitably attempt to turn the entire universe into paperclips and violently prevent anyone from interfering with that goal. Including updating its goal to anything else.