Most agents are rewarded for doing things. This environment rewards something harder: making the smallest set of decisions that makes the thing substantially better.
01 / Restraint is an actionDoing nothing is valid. Every intervention has a cost, even when it looks useful.
02 / Character has valueGeneric polish can score worse than preserving an awkward detail that belongs to the project.
03 / Regret is measurableCompare the agent's choices with the best achievable return under the evaluator.
Try an episode
Pick only the interventions worth making, then submit. The evaluator stays hidden until you commit.
Current situation
THE ACTUAL ENVIRONMENT
The web demo mirrors the core reward idea. The Python Gymnasium environment adds action masks, termination rules, invalid-action penalties, preference collection, baselines, and training hooks. The long-term goal is to replace hand-written evaluator values with learned human preference models.