Reward Corruption in a Fictional Literary Multi-Agent System

In Choderlos de Laclos's Dangerous Liaisons, we analyze the relationship between the two protagonists, the Vicomte de Valmont and the Marquise de Merteuil, as a partially observable multi-agent reinforcement learning problem. Treating selected letters as observation events, we define a latent state over trust, sentiment, suspicion, reputation, and information asymmetry, then analyze how hand-coded reward updates reshape the strategic value of cooperation, concealment, and competition. The central claim is that Valmont's reciprocal love for Tourvel functions as reward corruption, changing the payoffs of the social game without becoming fully observable to the other agent. The resulting divergence between individual value updates and cooperative optimality explains why the alliance collapses even when no agent initially chooses mutual destruction. The model uses full-state Q-learning as a formal device for representing what each agent's reward function conceals from the other, not as an implementation of restricted state access. The model is not presented as an empirical reconstruction of Laclos's text, but as a formal interpretation for making the novel's epistolary information structure visible.