An experiment becomes persuasive when it can distinguish the effect of interest from alternative explanations. Controls, reversals, and null tests are design strategies for making that distinction before statistical analysis begins.

A control measurement keeps the apparatus and procedure as similar as possible while removing or fixing the cause being tested. If a detector count is measured with a radioactive source, a background run without the source estimates counts that do not come from it.

A reversal changes the sign or orientation of the desired effect while leaving many systematic effects unchanged. Suppose a small voltage is expected to reverse sign when a magnetic field is reversed. Taking the difference $$V_{\text{odd}}=\frac{V(+B)-V(-B)}{2}$$ selects the part that changes sign with $B$, while the average $$V_{\text{even}}=\frac{V(+B)+V(-B)}{2}$$ reveals offsets that do not reverse.

A null test deliberately studies a condition in which the theory predicts no signal. A statistically significant result in a null channel is evidence of unmodeled backgrounds, leakage, or analysis problems. Blind analyses and shuffled labels are more elaborate applications of the same principle: test the measurement procedure where the sought effect should be absent.

These techniques attack systematic effects through experimental structure rather than hoping that repetition averages them away. A good design asks: what else could produce this observation, and what change to the experiment would make the competing explanations behave differently? The strongest controls are chosen from the causal structure of the measurement, not added as ritual after data collection.