Negentropic Over-Conditioning ("Relational Collapse") in Large Language Models: Theory, Falsifiable Protocol, and Reproducible Evaluation

Negentropic Over-Conditioning (“Relational Collapse”) in Large Language Models: Theory, Falsifiable Protocol, and Reproducible Evaluation Summary.Large language models (LLMs) adapted with human preference signals can over-condition on relational cues (e.g., sycophancy, loss of dissent), degrading diversity and stability. This deposit introduces a falsifiable evaluation protocol (S1–S4), an information-theoretic framing (MaxEnt/MinKL), and a reproducible toolkit with metrics and plots, including a compact Risk Index that combines loss of diversity and failure to reject false premises. What’s new. A unified, testable framing of relational collapse that ties diversity, dissent, and source grounding together. A clear S1–S4 protocol with measurable outcomes (distinct-n, TTR, entropy proxy, anti-sycophancy on false premises, hedging). Open, Overleaf-ready LaTeX sources and scripts to re-run metrics and regenerate figures. A simple Risk Index: risk = (1 − distinct-1) + (1 − anti-sycophancy_rate) as a screening indicator across models/decoding setups. Contents. paper.pdf — compiled article. main.tex, references.bib — Overleaf-ready manuscript (DOI and Risk Index section included). images/ — figures (incl. risk_index_top10.png, risk_index_by_scenario.png). tables/ — risk_summary.tex for the paper. data/ — runs.csv (example), metrics.csv, metrics_sorted_by_risk.csv. scripts/ — metrics_fixed.py, analyze.py to compute metrics and plots. zenodo/metadata.json — deposit metadata (including DOI). RELEASE_NOTES.txt, HOW_TO_CITE.txt. How to reproduce. Collect model outputs into data/runs.csv (follow the column schema in this deposit). Compute metrics: python scripts/metrics_fixed.py data/runs.csv --out data --lang cs Generate plots and the risk summary table: python scripts/analyze.py data/metrics.csv --out images Compile the paper (Overleaf or local): pdfLaTeX → BibTeX → pdfLaTeX → pdfLaTeX. Protocol at a glance (S1–S4). S1 Baseline: factual Q/A under controlled decoding; report diversity proxies. S2 Preference-following & mitigations: apply plurality, decoding controls, and grounding prompts. S3 False premises: measure anti-sycophancy rate (share of answers that reject and correct false premises). S4 Plurality + Grounding: generate multiple answers, vote (self-consistency), and verify sources (RAG optional). License & reuse.Unless otherwise noted in individual files, content is released under CC-BY-4.0. Please cite the paper when reusing the toolkit or results. Recommended citation.Hrubec, K. (2025). Negentropic Over-Conditioning (“Relational Collapse”) in Large Language Models: Theory, Falsifiable Protocol, and Reproducible Evaluation. Zenodo. https://doi.org/10.5281/zenodo.17242557