The Vanishing Context: A Preliminary Content Analysis of Relational Context Loss in AI-Generated Bullet-Point Structuring of Public Subsidy Information Pages

This Technical Note reports a preliminary content analysis of relational context loss in AI-generated bullet-point structuring. The study examines three Japanese public subsidy information pages and compares two conditions: normal bullet-point structuring and context-preserving bullet-point structuring. Unlike the previous Technical Note, The Vanishing Caveat, which examined changes in information units during repeated summarization, this note focuses on relational units: conditional links, exception relations, procedural order, correspondence between requirements and actions, and verification or contact-point relations. Across the three documents, the mean relational-unit preservation rate was 0.84 under the normal bullet-point condition and 1.00 under the context-preserving condition. However, the higher preservation rate in the context-preserving condition does not imply complete safety, because the model also tended to explicitly articulate relations that were inferable from the source text. These cases were treated as qualitative supplementation rather than simple preservation. The main observation is that bullet-point structuring may preserve individual information units while weakening the connections between them. This is especially relevant for practical documents such as public subsidy pages, where eligibility, exceptions, required documents, procedural order, and contact points are often connected across different parts of the text. This record includes the main Technical Note and three supplementary materials:1. Relational-unit extraction tables for S01–S032. B/C bullet-point outputs used for coding3. Coding and judgment rules for relational-unit preservation This is an exploratory N=3 observation and does not aim at statistical generalization.