Falsification tests distinguish frequency-aligned separation from α-lactose-specific detection in processed continuous-wave terahertz pellet spectra

Abstract Low-amplitude features near known absorption bands may be interpreted as analyte detection, although pellet and reference effects can generate similar structures. Using a previously described continuous-wave terahertz platform, we prepared α-lactose-containing and matrix-only control pellets and acquired separate datasets. Two separately prepared matrix-only controls exchanged reference and held-out comparator roles. Two independently prepared nominal 0.1 wt% pellets were evaluated at the pellet level, with repeated scans treated as nested technical measurements. Template and training-derived spectral consistency criteria were established from 0.5–8 wt% concentration-level means without the 0.1 wt% spectra. Under a constrained matrix-reference scaling correction, both 0.1 wt% pellet endpoints at 0.532 THz exceeded their fold-specific held-out control endpoints under both reciprocal reference roles, and their centers lay within the higher-concentration frequency range. However, the residuals failed the locked shape and width criteria, did not dominate matched off-peak responses, persisted after target-region masking, and were not jointly sign-stable under an unconstrained affine adjustment. These results support frequency-aligned spectral separation across two independent preparations under constrained comparison, but not α-lactose-specific detection, formally established limits of detection or quantification , validated quantification, population-level reproducibility, or generalizability. This study shows how post-acquisition falsification analysis can retain positive spectral evidence while bounding stronger analytical assertions.