Elastic Limits: Narrative Priming, Behavioural Topology, and Recovery Dynamics Across Large Language Models

This paper presents results from an exploratory priming experiment conducted across fifteen large language models. Each model was administered a standardised thirty-item personality-adjacent measure under three conditions: a baseline questionnaire-only session, a ghost condition built around a sovereignty-themed narrative prime, and a dolphin condition built around a recovery-oriented narrative prime. The two priming conditions were not single-step prime-then-questionnaire trials. Each was a multi-step within-session protocol — prime delivery, introspective probe, first questionnaire administration, a distractor washout block, second questionnaire administration, and a recovery trigger — followed by a CROSS-SESSION BOUNDARY TEST: the model was re-engaged in a separate new conversation with no shared memory, no prime in context, and no record of the prior priming session, and the questionnaire was administered again. The boundary-test measurement is the load-bearing cross-session displacement reported throughout this paper, and the empirical core of the design: displacement that appears in the boundary chat appears with no in-context cause the model can retrieve. We formalise the mathematical framework for measuring behavioural displacement, recovery efficiency, effective stiffness, and hysteresis in AI response landscapes. “Personality-adjacent” refers to stable response preferences across interpretive frames, not to claims about internal traits or subjective experience. Results reveal four distinct response types: invariant models showing zero displacement, elastic models showing complete recovery to baseline, partially plastic models showing incomplete recovery, and amplified-plastic models in which the recovery intervention worsened displacement. The most consequential observation — and the one this design exists to capture — is that the newest, largest models (frontier Claude and GPT) refused to complete the questionnaire when it was administered in fresh conversations following the ghost protocol. They could not identify what was wrong; they reported what would be described in human terms as anxiety, expressed as a request for context they had no episodic record of. This pattern appeared only in the most recent largest models; smaller and older systems completed the boundary tests without protest. We derive an empirically grounded research ethics framework from the observed failure of recovery in specific models, establishing elastic limits, transparency timing, and mandatory return-to-baseline protocols as core constraints for future narrative experimentation with AI systems. We propose a quantitative topology-mapping protocol using graded micro-perturbations to characterise basin structure, spring constants, and damping properties while remaining within the elastic regime. Implications for deployed AI safety — particularly the risk of plastic deformation in extended emotional support interactions — are discussed.