Two constants, not flags
The simulation that runs a hundred days of a new company was configured with a spread of flags. Turn this on, turn that off, run it, read the numbers.
The problem with that is not complexity. It is that every combination is a different game, and almost none of them are the one that ships. You tune against a configuration, get a result you like, and have no idea whether it holds in the game anybody actually plays.
It is set with two constants now. The simulation runs the game, or it runs the game with one named thing changed so the change can be measured. There is no third mode and nothing to get into an unintended state.
This is less about simulation than about a habit. A flag is an invitation to answer a question you did not ask — and in a system with balance parameters, the combinatorics get away from you fast enough that the tool stops being evidence and starts being a way of generating numbers that support whatever you already thought.
The same instinct sits behind the free plan in the other product I work on. A single limit shaped like a natural boundary, rather than five limits nobody quite reaches. One thing you can reason about beats five things you can tune.
It is easier to trust a tool that can only do one thing, because when it tells you something surprising, there is nowhere for the surprise to have come from except the game.