AI ASSISTANCE FOR PREDICTING NARRATIVE OUTCOMES

Abstract Psychodynamic clinicians face a persistent challenge: their claims are typically validated by postdiction — they apply concepts to narratives, such as case histories, whose outcome is already known. They rarely offer prospective predictions. This paper presents a systematic, rule-based framework for predicting the outcomes of stories of human actions. If these findings are valid, the methods used may be useful for similar attempts to leverage LLM strengths in further research.This framework, "L100-500," has five levels of specific rules. (See Appendix A, List of rules in L100-500 Ver. 4.2.) The author developed these rules through an iterative process—predict, test, revise, and re-test—using LLMs. When rules failed to predict a narrative outcome, the author revised them. Method: The project began by analyzing 14 New Yorker short stories published after August 2025 and therefore unlikely to have been included in the principal LLM's training data. That initial corpus was used to train Version 1 of L100-500. Version 1 rules were then applied to additional narratives drawn from television scripts, feature-film scripts, psychoanalytic case histories, mythology, and clinical supervision summaries. When Version 1 failed to reach adequate accuracy (better than 85%), Version 2 rules were developed to address its shortcomings. The same occurred with the development of Versions 3 and 4. Finally, when Version 4 produced better than 85% accuracy, it was codified as L100-500, Version 4.2.This codified rule set, L100-500, was then applied to 43 new texts drawn from the same kinds of narratives named above, along with classical literary works from non-Western traditions. This testing yielded accuracy estimates ranging from 85% to 92%. For each narrative, the opening 10% of the text was provided to an LLM, which was instructed to ignore prior knowledge of the text, to apply the L100-500 rules, and to generate falsifiable predictions about the trajectory of the main characters. The LLM then received the full text and was asked to grade the accuracy of its predictions.These findings are preliminary. Independent judges have not yet validated these high scores and the framework's apparent utility.