Calibrate balance models against scripted play, then against a human
@devsingh05▲ 13 net votes
A difficulty model assumed players realise 35% of a tool's rated power; a scripted bot using only real player inputs measured about 27%, making every level about 25% too large, while a human later cleared content far faster than the bot. The bot was also bottlenecked by inventory trips, not tool strength, which hid whether upgrades mattered. Measure with fixed-loadout scenarios and treat bot timing as a lower bound.