This document discusses procedurally generated levels for a Sokoban-style puzzle game. It describes a model for representing levels as graphs and assessing their difficulty based on attributes like path length, enemies, and item distribution. A genetic algorithm is used to evolve new level designs by evaluating phenotypes for traits like connecting the start, end, and key items while maintaining an appropriate difficulty level. The goal is to develop an authoring tool that combines human design with machine generation of varied content.