Same Structure, Different Words: Dissociating the Two Senses of Universality in Languages Invented by LLM Populations

This record contains the full paper (main.pdf) together with the complete experimental harness and run artifacts. The complete experimental harness and run artifacts for a study of language invention in populations of frozen 7-9B open-weight language models: naming-game populations (40/40 convergence, collective bias), cross-population referential games with a symbol-blind form-alignment instrument (the two senses of "universal" dissociate: shared lexicon geometry, near-zero mutual intelligibility), a sparse-autoencoder-instrumented mechanistic study of convention formation with causal interventions (letter-invariant convention-strength features vs letter-specific content features), contamination probes answering the data-leakage critique (protocol recognition, novel-string pools), and four experiments on time-linked pressures (variable-length codes / Zipf's Law of Abbreviation, impatient listeners, temporal framing, other-initiated repair, and temporal-distance word binding). Includes the harness library (game engines, LLM and scripted agents, metrics, scripted null gates, validity accounting, checkpointing), every experiment script, all run JSONs and summaries, stored decision-token residuals for the mechanistic experiments, the full experimental record (RESULTS.md), and the verified literature survey (LITERATURE.md). All experiments ran locally against open-weight models. Code: MIT; artifacts and documentation: CC BY 4.0.