Linguistic Polarity and Decision Architecture in LLM-Based Abstract Screening for Systematic Reviews

Large language models (LLMs) are increasingly investigated for abstract screening in systematic reviews, yet it remains unclear whether screening errors attributed to linguistic complexity arise from intrinsic semantic sensitivity or from its interaction with decision architecture. We examined how five polarity variants of logically equivalent eligibility criteria—affirmative inclusion, antonymic exclusion, predicate negation, verb-level negation, and double negation—affect screening outcomes in a controlled biomedical task. Using 1000 abstracts from a reconstructed Cochrane review corpus (50 TARGET; 950 non-targets), we implemented four abstract-visible criteria within a sequential hard-gated pipeline, where failure at any step triggered irreversible exclusion. Under hard gating, linguistic polarity alone produced substantial and statistically significant variation in recall. For GPT-5.1, recall ranged from 0.72 to 0.32 despite identical logical predicates and input data. Replication with GPT-3.5 Turbo yielded a similar divergence (0.92–0.18), confirming generalization across model generations. TARGET losses were concentrated at criteria typically satisfied but inconsistently reported in abstracts, indicating conservative exclusion under evidential under-specification. To assess whether this effect is semantic or architectural, we reimplemented screening using a scoring-based evidence-accumulation framework in which each criterion contributed graded support and inclusion was determined by a tunable threshold. Under scoring, recall increased across all variants and converged within a high-sensitivity regime, while residual polarity effects were attenuated but remained detectable. Linguistic differences shifted from structural recall collapse to controlled precision–recall trade-offs. These findings show that negation sensitivity is strongly mediated by decision architecture. Irreversible gating amplifies local uncertainty into false-negative exclusion, whereas cumulative scoring preserves uncertainty and enables controllable operating thresholds.