The End of Defaults: A Compositional Shift in Public Telegram Channels in Four Countries, 2016–2025

A preprint and reproducibility package on how the composition of public feeds changed between 2016 and 2025. Paper — A Compositional Shift in the Feed. 2,868,643 messages from 59 public Telegram channels in Russia, Ukraine, the United States and Brazil, labelled against ten themes grouped into macro-thematic and locally discrete classes. Within the same channels the macro-thematic share rose in all four countries, survived the removal of the windows around 23 pre-registered shocks, and reached meme and lifestyle channels; a matching within-channel fall of local themes is established only in Russia. The common date is 2022, not 2020. Two pre-registered hypotheses were refuted: the feed did not accelerate (weekly vocabulary turnover fell), and feed heaviness relates to epochal phrasing positively, not negatively. A direct test of whether the theme classes are a difference of form (events versus states of affairs) did not confirm it: the paper claims a change of composition, not of form. Essay — The End of Defaults. Twelve spreads reading the same measurements against a model of the decade, with every claim graded: measured, established, compatible but not shown, not established. Forecast — The Closable Unit. A predictive essay on interface design with nine registered, falsifiable forecasts resolving between 2029 and 2033. Registry v6 (FORECASTS_v6.csv, sha256 17f60a10a745a3641244a2b58ff8c184a00748d5685297381975f53e9032818b) and its day-zero baseline were measured under an audit protocol sealed before its first query. This record is the public timestamp of that registry. Package. Preregistrations and amendments, analysis scripts, all result files, item-level labels (without message texts), the channel roster, the forecast registry with every version and audit capture, and verify_package.py, which checks the package against itself. Use of AI. Item-level coding was performed by two large language models (Claude, Anthropic; MiMo, Xiaomi). Claude was also used to write analysis code and draft the text under the author's direction. Every number is read from the published result files; the author takes full responsibility for the content. Message texts are not published: they are other people's messages. Russian editions are included; Russian is the lead language of the essays.