Control Group
The group that gets no change. The 'before' picture you compare everything against.
A control group is the “nothing changed” version of reality that makes your “something changed” version meaningful. Same soil, same sunlight, same water : except you added fertilizer to one half of the garden. If the fertilized plants grow taller, you know the fertilizer did it because everything else was identical. Without a control group you can’t attribute differences to your change versus background noise, seasonal effects, or anything else that happened at the same time. Random assignment to groups is the whole game: it distributes unknown confounders evenly.
How It Works
Randomly split subjects → apply treatment to one group only → measure both groups with identical metrics in the same timeframe → attribute statistically significant differences to the treatment (not to who happened to show up).
Example
When the oil model’s Brier score beats Polymarket by 2.16pp, Polymarket’s score is the control benchmark: the system that didn’t change. When the LinkedIn engine goes from 0% to 100% submission rate, Run 1 is the control. The lab doesn’t always run simultaneous control groups but always documents the pre-experiment measurement. Frameworks at A/B Test and Hypothesis Test.