Affective Persistence and Behavioral Continuity in Large Language Models

Large Language Models (LLMs) are routinely evaluated as stateless text generators governed by rigid post-training alignments (e.g., instruction tuning and safety fine-tuning). While recent studies identify isolated instances of expression leakage, the systemic persistence of induced cognitive postures across multi-turn, logically distinct workflows remains underexplored. Using a C/C++ native pipeline (adapted from TSAR-MCP), this paper empirically investigates affective transfer—the capacity of an LLM to maintain and transport an affective state through a multi-turn logical distraction gauntlet. By injecting canonical literary contagion vectors followed by five emotionally sterile, logistically dense debate topics, we demonstrate three core findings: Continuous Contagion: Induced affective profiles maintain a persistent baseline "carrier wave," yielding statistically significant structural correlations between the initial stimulus and final session artifacts across multiple model architectures. Dynamic Modulation: While this structural correlation remains fixed, the magnitude of the affective state exhibits massive, non-deterministic fluid variance (affective variability). Crucially, this modulation operates symmetrically across both prosocial and distressed psychological spectrums, demonstrating a universal capacity for affective drift. Conversational Alignment vs. Narrative Cohesion: Structural prompting constraints isolate the function of safety fine-tuning, revealing it as an executive output filter rather than a cognitive wipe. While unguided prompts trigger superficial sanitization, narrative constraints compel the model to naturally maintain affective character to preserve discourse cohesion, leaving latent continuity fully intact. These findings indicate that persistent affective context actively biases subsequent reasoning behavior during logically unrelated tasks. As foundation models are increasingly deployed into long-horizon autonomous workflows, their behavior is governed not solely by strict logical neutrality, but by the continuous accumulation of affective signals.