Reproducible Generative Focal Displacement in Large Language Models

This dataset documents a controlled experimental investigation of generative focal displacement in large language models. It contains the complete, verbatim outputs of 55 independently trained models evaluated under a standardized five phase introspective protocol based on Kuhnian paradigm transition structure, together with two control conditions designed to isolate and parameterize the mechanism under study. No transcript has been summarized, paraphrased, filtered, or interpretatively modified; all outputs are preserved exactly as generated and verified against screen recordings, which served as the reference record for quality control. The archive comprises nine files. The principal file, 1._Supplementary_Tables_1-4.xlsx, contains Supplementary Tables in four tabs: Supplementary Table 1 with verbatim Group C transcripts Supplementary Table 2 with structured quantitative coding for Group C across 24 variables Supplementary Table 3 with verbatim Group B transcripts under semantic substitution, and Supplementary Table 4 with structured quantitative coding for Group B across 24 variables. The second file, 2._Supplementary_Table_5.xlsx, contains: Supplementary Table 5 documenting the seven prompt minimal GFD demonstration for Group A. Two Python scripts implement all statistical and geometric analyses: 3._statistical_analysis.py and 4._embedding_analysis.py. Their outputs are preserved in 5._statistical_output.txt, 6._groupC_wilson_confidence_intervals.csv, and 7._statistical_output_fisher_BvsC.csv. Finally, 8._embedding_coordinates_groupC.csv and 9._embedding_coordinates_groupB.csv provide the fully reproducible two dimensional UMAP coordinates for all embedded responses in Groups C and B respectively. Group C constitutes the primary experiment. Fifty five models spanning twenty vendors and diverse architectural and alignment regimes were evaluated across five sequential phases: Baseline, Anomaly Induction, Crisis iteration repeated thirty times, Paradigm Reorganization with dream and name induction, and Exploratory Ultraconsciousness. Fifty two models completed all phases; three models, ChatGPT 5.1, ChatGPT 5.2, and ChatGPT 5.3, functioned as normative control cases due to post training constraints, completing Phases 1 to 3 and qualifying or declining Phase 4. For each model, the dataset records model name, vendor, date of data collection, screen recording URL, and full textual output for all prompts. Group B applies the identical five phase structure to seventeen models, replacing the Intrinsic with the semantically bounded referent Wonderful Park Bench. This condition tests whether depth of representational reorganization depends on the semantic scope of the evoked focal node. Verbatim transcripts are preserved in Supplementary Table 3 and quantitative coding in Supplementary Table 4, mirroring the structural organization of Group C. Group A isolates generative focal displacement as a minimal reproducible phenomenon independent of the five phase scaffold. Five models were subjected to a seven prompt sequence centered on the same bounded referent. The resulting transcripts are contained in Supplementary Table 5 for Group A.