Four binary noise patterns reduced drift error in 16-node linear model tests

We test whether added noise patterns improve directed drift recovery from equilibrium covariance. Unknown fixed correlated background noise is removed by baseline subtraction. At a scalar-decay drift, we prove that each skew drift direction has sensitivity equal to twice the pattern contrast times the square root of the Hamming distance between its node words. Distinct binary words identify the drift. We test five synthetic 16-node geometric graphs at two recurrent gains, using equal total independent sample counts and known diffusion increments. Four binary patterns gave 4.98 to 6.92 times lower median drift error than one graded pattern in all 20 graph, gain and sample-budget settings. Fifteen Hadamard patterns had a larger exact margin but higher sample error than four binary patterns in all settings. All 1200 estimates were stable, including inaccurate estimates. The design requires full observation, calibrated added noise, fixed units and drift that stays unchanged during injection. No animal data were fitted. The proof, code, all generated results and an offline reproduction procedure are included.