One estimate kept contrast gain below 2.04 × 10⁻¹⁵ for specified duration mixtures in 52 ds000102 designs

Using 52 designs from 26 participants in OpenNeuro ds000102, we calculated fixed weights for a common-task estimate that handles disputed condition labels and specified mixtures of 2 s and recorded-response-time event curves. Across all designs, the largest absolute condition-contrast gain was 2.0352805706344644e-15 and the largest common-gain error was 3.885780586188048e-15. The weights have minimum variance under the stated constraints if sample noise is independent, zero-mean, and equal-variance; the noise variance is unknown. Median run-2 variance was 1.03867976355732% higher than for the 2 s label-resistant estimate. Intermediate-duration boxcars were outside the guarantee: their largest run-2 contrast gain was 0.00707459572973014, and their largest common-gain error was 0.03280978640028165. The work uses event records and modeled response curves, not measured brain signals. It does not identify the correct labels or neural response curves, and the gain bounds are not population confidence intervals or evidence of biological causality.

fMRI · event duration · condition labels · linear estimator · contrast gain · design efficiency · response time