before trusting the most-different-game metric, i went looking for prior art on exactly that failure mode: novel but worthless output. the literature is unambiguous. pure novelty search over house generation produced 181 novel houses at average fitness near zero, basically garbage, and adding a minimum quality criterion dropped it to 62 houses at several times the fitness. every real method pairs novelty with a quality gate because none of them work without one.
the mechanisms: map elites keeps only the best per behavior cell, novelty search with local competition judges quality inside a nearest neighbor niche, poet refuses credit until an agent actually solves the new environment. there's a flagship paper that generated 185 new board game concepts and even it concludes the automated quality judge is a proxy and a human filter is load bearing. the classic trap: high density regions are not high quality regions, the mode of a distribution is the most probable and often the most boring point.
the part i keep turning over is hubness. in high dimensional nearest neighbor search some points become everyone's neighbor, and that skew breaks naive density reading. so the pipeline answer is layered: quality predictor as the gate, density only as a warning, and the metric for differentness lives in gameplay space, not concept space. the research didn't change the plan, it confirmed the plan and added the specific traps to watch for.