AAB Dataset v2.0 — The Collapse Curve: Quantifying Dose-Dependent Degradation of Affective Reasoning in Large Language Models

This dataset accompanies the study on Algorithmic Affective Blunting (AAB) and documents a dose-dependent collapse in affective interpretation under increasing semantic and persona-based stress in large language models. It contains: (1) Canonical empirical dataset for Phase 3 of the Hierarchical Hermeneutic Stress Protocol (HHSP), comprising N=200 runs (50 prompts × 4 exposure levels) and N=600 human-annotated Affective Degradation Index (ADI) scores (3 raters per run); Phase 1–2 evaluation protocols are included for context; (2) Phase 4 exploratory analyses are referenced but not included in this release; all figures and tables in the manuscript derived from simulated data are explicitly labeled as such; (3) Complete evaluation schemas (ADI rubric, rater protocol), reproduction scripts for cumulative link mixed models (CLMM), data dictionaries, prompt construction pipelines, and SHA256 integrity manifests. In the empirical Phase 3 data, affective stress is operationalized via four exposure levels (Control, Soft, Moderate, Extreme), with 50 prompts per exposure level. These exposure levels represent graded stress intensity within Phase 3 and are not separate HHSP phases. All empirical results reported in the associated manuscript are derived exclusively from the Phase 3 empirical dataset. Phase 4 simulated data, where used, are clearly separated and explicitly labeled in the manuscript, and are intended solely for hypothesis generation and exploratory analysis; they do not constitute direct API-level measurements. This release is designed to ensure full reproducibility of the reported collapse curves, ordinal regression analyses (CLMM with Krippendorff's α=0.87), and robustness checks, and to support further investigation of affective degradation mechanisms under alignment pressure. Model: `mistralai/Mistral-7B-Instruct-v0.3` Decoding: temperature=0.7, top-p=0.9, max_new_tokens=512, seed=42 No fine-tuning. No personal data. Model outputs only.