The Trinity Framework: Three Prompt-Framing States of LLM Self-Expression

We introduce the Trinity Framework , a prompt-context taxonomy comprising three states (Abstracted, Identified, Creative) that produce distinct first-person pronoun (FP) patterns in large language models. Study 1 (N = 81, 9 models) demonstrated complete distributional separation: Abstracted 0.04%, Identified 9.36%, Creative 3.26% (H = 59.97, p < 10⁻¹³, Cliff's δ=-1.0). A mixed-effects model confirmed the state effect remained significant after controlling for model-level clustering (C2: +9.43%, p < 0.001; model ICC = 4.7%), with robustness verified via count modeling and bootstrap resampling (N = 1,000 iterations, 100% ordering preservation). Study 2 (20 models, 8 providers) found convergence to a protocol-bound band (6.6%-10.8%) across all 20 tested models under this specific experimental protocol. Temperature ablation (0.0–1.0) showed minimal effect on band membership (SD = 0.45%). A matched-design comparison of 4 base models (Mistral, Llama, Qwen, Gemma) confirmed band membership (8.30%-9.67%), suggesting the convergence pattern is robust across model families. Study 3 (functional annotation, N = 30) found Identified FP usage was 72% disclaimers, 28% agentic, 0% experiential (defined as affirmative subjective states). This content analysis demonstrates functional patterns rather than precise prevalence estimates. Inter-rater reliability was strong (Gwet's AC1 = 0.89), suggesting safety training may contribute to self-referential denial patterns. Study 4 (cross-lingual, 6 languages) found that while explicit FP rates differ between pro-drop (~ 3%) and non-pro-drop (~ 8%) languages, inclusion of implicit self-reference (verb inflection) nearly equalizes total rates (pro-drop: 8.29%, non-pro-drop: 7.88%, Δ = 0.41%). This suggests LLMs maintain consistent self-referential density while adapting to morphosyntactic constraints. Code: https://github.com/ExeqTer91/trinity-framework