Surprisal maps differently onto online measures of sentence processing

Contextual predictability is a robust determinant of online sentence processing, but it is unclear whether predictability affects different online measures through a shared functional mapping. We address this question by comparing how surprisal relates to processing cost across multiple behavioral and electrophysiological measures in two complementary datasets: the Dutch RaCCooNS co-registered eye-tracking/EEG corpus and an English 205-sentence benchmark linking eye-movement, self-paced reading, and ERP data. Across both corpora, surprisal reliably predicted online processing cost. We then tested whether the same surprisal-to-cost mapping generalized across measures by estimating measure-specific power-law transformations under blocked cross-validation and confirmatory full-data models. The results did not support a single invariant linking function. Instead, the preferred mapping varied across measures within each corpus. In Dutch natural reading, first fixation was closest to ordinary surprisal, whereas later eye-movement measures and the fixation-related N400 favored steeper mappings. In the English benchmark, reading-time measures clustered in a compressed regime, the N400 was sublinear, and the P600 was superlinear. A targeted GAMM robustness analysis supported the main within-corpus dissociations. These findings argue against treating major online measures as interchangeable readouts of a single surprisal-linked cost signal. Surprisal remains central to sentence processing, but the mapping from surprisal to observable cost is itself part of the psycholinguistic theory that needs to be explained.