Synthetic Worldview Reconstruction (SWR): An LLM-Supported Method for Analyzing Latent Belief Systems

Abstract This work introduces Synthetic Worldview Reconstruction (SWR), a new social science method for reconstructing and comparing the collective worldviews of sociologically defined groups from publicly available data. SWR exploits the synthetic integration capability of Large Language Models (LLMs) to derive a group's latent belief system from the aggregated public statements and documented actions of its members. The method operationalizes this core idea through a standardized five-step prompt protocol, systematic data grouping into synthesis units, and a multi-layered validation framework. SWR was developed and empirically hardened in a large-scale application study (DOI: 10.5281/zenodo.18736737) and is presented here as a generalized, transferable framework. Changes in v5.0 (Consistency Patch) Aligned with Paper A v7.0: All empirical validation results updated to match the current state of the companion application study. IMIIRR introduced: Intra-Model Inter-Instance Rating Reliability (N=15 independent single-instance runs, ICC(3,1)=0.847, Mean SD=0.45) documented as psychometric property of SWR. Instance separation: Established as critical methodological requirement throughout both language versions. Validation framework restructured: V1 changed from Inter-LLM Reliability to Instrument Calibration (Run Convergence). Inter-model comparison moved to Tier 2 (desideratum). Cross-modal prediction corrected: 88% to 72% confirmed (matching Paper A v7.0 data). Simpson's Paradox: Group-level say-do decomposition findings added to pilot section. 12 QA measures: Reference to expanded quality assurance regime from Paper A v7.0. Narrative corrected: Paper A developed the method, Paper B generalizes it as transferable framework. GitHub URL updated: research-line/ai-elite-swr (DE version had outdated URL). Group-level framing: Corrected LLM-as-Participant comparison to reflect SWR's primary group-level operation. Changes in v4.0 Paradigm shift to group-level analysis (matching Paper A v6.0). Terminology: "Synthesis Units" replaces individual-focused framing. Self-contained methodology (no companion paper required for comprehension). Four theoretical anchors: Weber's ideal type (primary), Booth's implied author, IRL, projective tests. Epistemic privacy section added. PersonaCite (RAG-based persona) comparison added. Two versions: English (32 pages) and German (34 pages). Note: This preprint has not yet undergone formal peer review. v5.0 is a consistency patch aligning Paper B with the empirical findings of Paper A v7.0. A comprehensive rewrite of Paper B is planned as a future step.