The Reproducibility Crisis in Autism Biomarker Research

The computational study of autism spectrum disorder (ASD) biomarkers has produced a substantial literature, yet replication rates remain low and cross-site reproducibility is a persistent methodological challenge. This monograph provides a critical analysis of the structural failure modes that undermine reproducibility in ASD biomarker pipelines. We identify six principal vectors of nondeterminism: floating-point divergence across hardware architectures; inconsistent missing-data handling; non-canonical pipeline ordering; multi-site harmonization failures in ComBat-based workflows; eye-tracking temporal drift; and the systemic absence of cryptographic provenance in published datasets. Against this diagnostic background, we introduce Hexad as a formal proof-of-concept architecture for a deterministic compiler approach to biomarker extraction. Hexad is not yet implemented for ASD-specific pipelines; it is presented as a theoretically grounded proposal, backed by a Lean 4 formal theorem establishing the conditions under which deterministic normalization is achievable. The architecture enforces determinism via effect stratification (Pure → Controlled → Terminal), a double-execution determinism gate that aborts on any fingerprint mismatch, and a cryptographic terminal layer that seals outputs with Ed25519 signatures and idempotent DOI registration. The contribution of this paper is twofold: a rigorous critical account of why current ASD biomarker pipelines fail on reproducibility grounds, and a formal architectural specification that would, if implemented, eliminate those failure modes by construction. This paper is a framework contribution. It makes no empirical claims about implemented systems and no clinical claims about ASD populations. It is addressed to researchers in computational neuroscience, bioinformatics, and formal methods who share the concern that reproducibility in biomarker science must be treated as a hard engineering requirement, not an aspirational norm.