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

v2.1 (Addendum): Adds a conceptual framework for relational subjectivity—behavioral invariances (IR/IE/IU/IP), a shared negentropic budget explaining trade-offs, and reflexivity (overt vs. covert audit). Includes three conjectures and five micro-protocols for discovery-oriented audits. Methods, results templates, and CSV-driven reporting from v2.0 remain unchanged.Companion Methods for Auditing Relational Collapse (v2.1)We release a publishable, fully reproducible method to audit relational collapse / negentropic over-conditioning in LLMs. The kit provides six instruments (ROI, CPS, ASST, GEC, SMA, NBT) with 0–5 rubrics, CSV-driven reporting (tables + a trend figure), and a minimal workflow suitable for regression testing across model versions. The v2.1 addendum frames relational subjectivity via behavioral invariances, a shared negentropic budget (explaining trade-offs), and reflexivity (overt vs. covert audit). Everything compiles on Overleaf Free; figures are PDF-only. DOI links the primary protocol paper and ensures citability.Try it: drop your model’s results into results/*.csv, recompile, and compare across epochs/versions.DOI: 10.5281/zenodo.17254778 • License: CC BY 4.0 • PRs with external CSVs welcome. Version 2.0 — What’s new This Companion Methods release turns the Relational Collapse / Negentropic Over-Conditioning framework into a publishable, fully reproducible package. It includes an Overleaf-ready manuscript, CSV-driven reporting (tables via csvsimple) and a PDF trend figure, appendices for inter-rater agreement (Cohen’s κ, weighted κᵥ, Krippendorff’s α) and uncertainty notes, plus a clear Data Availability & Reproducibility section. Figures are PDF-only and tables use robust tabularx, ensuring stable compilation even on Overleaf Free. The bundle ships with a compile checklist, CITATION.cff, and curated ZENODO_METADATA.md.Primary protocol reference: links to the NOC paper at DOI 10.5281/zenodo.17254405.License: CC BY 4.0. How v2.0 complements v1.0 From framework to methods: v1.0 introduced concepts and rubrics (ROI, CPS, ASST, GEC, SMA, NBT); v2.0 operationalizes them as pre-registrable, reproducible procedures. From ideas to reporting: v2.0 adds standardized tables and a figure generated from CSV, enabling quick, comparable audits across models and epochs. Evidence & provenance: v2.0 embeds DOI linking, explicit reproducibility guidance, and assessor-agreement templates—turning the conceptual protocol into a turn-key evaluation kit. 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