Hexad: A Deterministic Replay and Proof-Carrying Reproducibility Pipeline for Multi‑Site Biological Sensor Studies

Abstract The reproducibility crisis in computational biology, particularly in multi‑site autism spectrum disorder (ASD) biomarker research, stems from hidden nondeterminism in sensor acquisition, time‑stamping, and execution environments. We introduce Hexad, a deterministic replay and proof‑carrying reproducibility pipeline that guarantees deterministic replay with cryptographically attested trace equivalence and bounded cross‑site sensor drift. Hexad combines (1) a TLA⁺ specification of deterministic replay, (2) a Lean 4 machine‑checked verification framework with completed safety proofs and ongoing refinement proofs linking the implementation to the formal model, (3) a hardened C++ interposition layer with an mmap ring buffer and a determinism gate that aborts on any drift exceeding a calibrated dimensionless admissibility bound τ_drift = 0.03114, and (4) a cryptographic attestation that produces an Ed25519‑signed reproducibility report. We demonstrate the pipeline on interleaved eye‑tracking, EEG, and behavioural data, showing deterministic replay across two independent Linux workstations. Hexad provides a regulator‑grade, falsifiable, and verifiable reproducibility substrate for multi‑site clinical sensor studies and high‑stakes computational workflows, designed to support documentation and robustness requirements associated with EU AI Act Article 15.