How a shared state is described determines whether AI agents synchronize: data and code
Year: 2026
Type: software
Abstract: Complete research compendium for the study of how the encoding of a physical state into model input changes both the microscopic response operator of a language-model agent and the collective outcome of a population of such agents. The deposit contains the deterministic circular-agent engine, the observation-map encoders, the prompt contract, the acquisition runners with their versioned and hash-locked protocol files, the statistical analyses, a byte-for-byte frozen copy of every analysis artefact the paper reports, and the complete raw acquisition traces for all fifteen acquisition groups (one JSON line per model call: task and field identifiers, the physical hash of the field, the encoding, the acquisition block, the sampling seed, the SHA-256 of the prompt, the raw response, the parser stage that read it, token counts and elapsed time). No API key or access to a provider service is needed to reproduce the deterministic engine, the statistical analyses or any reported number. Exact re-acquisition depends on the continued availability of the specific model checkpoints. The figure sources and the rendered display items are included, so every figure in the paper rebuilds from the deposited artefacts.
Authors: Ezaki, Takahiro, Imura, Naoto, Nishinari, Katsuhiro
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7201848228