This paper explores the relationship between human sentence processing and grammatical knowledge, proposing a psycholinguistic model based on a probabilistic Earley parser. The author defines cognitive load in terms of the total probability of structural options that are disconfirmed at some point in a sentence, measured by the surprisal of a word given its prefix on a phrase-structural language model. The probabilistic Earley parser, developed by Stolcke, is used to efficiently calculate these probabilities and predict reading times word by word. Under grammatical assumptions supported by corpus-frequency data, the parser correctly predicts processing phenomena such as garden path structural ambiguity and subject/object relative asymmetry. The paper outlines three principles: strong competence, where the parser directly uses grammar rules; frequency affecting performance, where statistical theories of language performance are adopted; and eager processing, where the parser is not interrupted by information that could be used. The probabilistic Earley parser is shown to satisfy these principles by computing prefix probabilities, which measure the disconfirmation of structural analyses. The parser's ability to handle hierarchical phrase structure and its total parallelism in maintaining all compatible trees make it a powerful tool for modeling human sentence processing. The paper also discusses the plausibility of probabilistic context-free grammars, the garden-path effect, and subject/object asymmetry. It concludes that a "total-parallelism" parsing theory based on probabilistic grammar can explain important processing phenomena, suggesting that phrase-level contingent frequencies can account for many aspects of sentence comprehension.This paper explores the relationship between human sentence processing and grammatical knowledge, proposing a psycholinguistic model based on a probabilistic Earley parser. The author defines cognitive load in terms of the total probability of structural options that are disconfirmed at some point in a sentence, measured by the surprisal of a word given its prefix on a phrase-structural language model. The probabilistic Earley parser, developed by Stolcke, is used to efficiently calculate these probabilities and predict reading times word by word. Under grammatical assumptions supported by corpus-frequency data, the parser correctly predicts processing phenomena such as garden path structural ambiguity and subject/object relative asymmetry. The paper outlines three principles: strong competence, where the parser directly uses grammar rules; frequency affecting performance, where statistical theories of language performance are adopted; and eager processing, where the parser is not interrupted by information that could be used. The probabilistic Earley parser is shown to satisfy these principles by computing prefix probabilities, which measure the disconfirmation of structural analyses. The parser's ability to handle hierarchical phrase structure and its total parallelism in maintaining all compatible trees make it a powerful tool for modeling human sentence processing. The paper also discusses the plausibility of probabilistic context-free grammars, the garden-path effect, and subject/object asymmetry. It concludes that a "total-parallelism" parsing theory based on probabilistic grammar can explain important processing phenomena, suggesting that phrase-level contingent frequencies can account for many aspects of sentence comprehension.