Reproducible High-Throughput Simulation and Parameter Recovery for the Drift-Diffusion Model: The sinistra Library, with an Application to the Hand Laterality Judgement Task

Abstract

sinistra is a high-throughput drift-diffusion model (DDM) simulator implemented for reproducible parameter recovery, using a per-trial ChaCha8 random stream design that yields bit-identical results across CPU architectures, core counts, and thread counts — not just thread-count invariance. Benchmarked on arm64 (10-core) and x86-64 (2-core) machines, the simulator sustains over 1.4M trials/s and reproduces every non-timing result (parameter recovery, dt-scaling, and pooled/participant-level fits) to the last printed digit across platforms. Parameter recovery across 7 parameter sets (n=50,000, 5 reps) shows low bias for both EZ and simulation-based fitting; EZ's boundary-parameter bias is shown to scale with √dt, consistent with Euler–Maruyama discretization error. Applied to a hand laterality judgment task (N=40), the method finds that a drift-rate effect and a non-decision-time effect are comparably reliable (dz = −0.98 vs. +1.09), with non-decision time accounting for the majority of the observed reaction-time difference — refining an earlier drift-rate-only account of the effect.