Again, that's the beauty of the Sliding Window approach vs the Wild Guess Curve.

Assuming 51% of miners are honest and won't manufacture fake transactions just for the lulz, the limit will only increase if there's enough real demand for it.

This approach will also provide significant inertia to protect against occasional spikes. It also allows the limit to decrease if there's not enough demand.

I.e. it's adaptive, which is what you usually want in control systems. You wouldn't build an autopilot by hard-coding a guess of everything it will have to do in the air. Instead you will build it to listen to inputs and generate compensating actions.