Cross-Family Mimicry and Parameter Non-Uniqueness in Drift–Diffusion Inference: A Minimal Synthetic Stress Test

This technical note presents a controlled synthetic stress test on inferential non-uniqueness in drift–diffusion inference under finite behavioral interfaces. Using a minimal Drift–Diffusion Model (DDM) setup, the analysis examines two related questions. First, it tests whether parameter recovery remains stable when the same synthetic generative source is fitted through restricted observational channels based on accuracy only, correct-response reaction times only, and error-response reaction times only. Second, it tests whether a heterogeneous two-drift mixture process can be behaviorally mimicked by a simpler constant-drift DDM family under a common observable interface. The results are strictly local and diagnostic. They show that reduced observational channels can support divergent latent recoveries, and that low observable discrepancy can coexist with substantial latent structural difference across model families. The purpose of the note is not to introduce a new cognitive theory, nor to claim universal non-identifiability of sequential-sampling models. It isolates a minimal case in which apparently well-behaved behavioral observables fail to guarantee unique inversion toward latent decision architecture. The accompanying bundle includes the PDF, the standalone Python script, reproducibility notes, and archived reference tables used in the bundle. The code is provided as a minimal synthetic stress-test implementation designed to reproduce the logic and structural signature of the analysis rather than to serve as a full inference pipeline. This metadata wording also stays aligned with the program’s Zenodo policy of keeping records clean, local, readable, and methodologically disciplined.