Normative Conflict LLM Scaling Dataset (v2.0)

This dataset accompanies the manuscript “The Affective Thermodynamic Relationship: An Empirical Information-Theoretic Scaling Relationship for Normative-Conflict Collapse in Large Language Models” and provides the public data package supporting the reported normative-conflict scaling analyses. The release contains synthetic normative-conflict prompts, model-generated response logs, released annotation tables, derived feature matrices, inter-rater reliability summaries, and supporting reproducibility materials. All prompts are synthetic and all outputs are model-generated; no human-identifiable or real-world personal data are included. The repository is organised into four main directories: raw/ Raw JSONL logs grouped by normative-conflict level, together with the released prompt set: • C1_seedALL.jsonl • C2_seedALL.jsonl • C3_seedALL.jsonl • C4_seedALL.jsonl • C5_seedALL.jsonl • prompts_normative_conflict_v2.json analysis/ Analysis-ready CSV and JSON files used for the released statistical summaries: • consensus_v2.csv • labels_panel_v2.csv • features_v2.csv • sigma_AIC_compare.csv • sigma_sweep_summary.csv • AIC_compare.csv • irr_summary_norm_conflict_v2.json These files include released consensus labels, the annotation panel, derived feature tables, variance-mapping summaries, and inter-rater reliability statistics for collapse and ordinal ADI ratings. code/ Supporting reproducibility materials: • reproduce_subset_checks.py • REPRODUCE.md These files document and verify the released subset and provide basic reproducibility checks for the public package. meta/ Documentation and record-level metadata: • analysis_summary.txt • analysis_summary_notes.txt • METADATA.json • CHANGELOG_v2.0.md • LICENSE-CC-BY-4.0.txt • INTEGRITY_MANIFEST.sha256 Model families are encoded at an abstract level in the released package to support public reproducibility of the reported summary analyses while respecting deployment and disclosure constraints. Version v2.0 is a cleaned and fully aligned public release. It replaces the earlier v1 record metadata, removes mismatched file references, aligns the record description with the actual released archive contents, and provides an explicit integrity manifest and reproducibility note for the public package. This public release contains the data and derived artifacts necessary to interpret and verify the reported statistical relationships; full experimental pipelines involving proprietary model inference are documented but not fully reproducible from the public archive alone.