Design-Based Analysis of Regression Discontinuities: Evidence from an Experimental Benchmark