The correlation between acceptability and probability of a sentence is an important issue both for theoretical and computational linguistics. Previous studies on this topic compared acceptability judgments and probability scores based on the unambiguous linguistic phenomena. The fact whether language model can predict possible agreement variation has been ignored. This study presents an experiment of modeling agreement variation with probability metrics. The predicate agreement with the coordinated subject in Russian was observed. We compared the acceptability judgments passed by the language speakers and the probability metrics predicted by the language model ruBERT-base without fine-tuning. We analyzed in detail which syntactic, morphological and semantic factors influence the acceptability and the probability of the sentences. The results show that the BERT language model fails to predict the agreement strategy given the variation. The subject-predicate order turns out to be the only factor which equally influences the acceptability and the probability of the sentence. Thus, the language model without fine-tuning cannot be used to parametrize predicate agreement with the coordinated subject. The sentence acceptability is based on the subtle linguistic contrasts which are not significant for the computer evaluation of its probability.