Year: 2026
Type: preprint
Abstract: Supplementary artefacts for the paper "Falsifiable Theory of Mind: Pairing Multi-Agent LLM Reasoning with Deterministic Outcomes".This deposit contains the complete record of a single certified 2v2 multi-agent match run on the iXentBench Arena platform (session A-2v2_L1_Multi_20260410_205616, April 2026): 48 turns on a 4x4 board, four LLM agents from two model families (gemini-2.5-pro and gemini-3.1-pro-preview), terminated on move limit. Every figure reported in Section 5 of the paper is reproducible from these files without access to the platform.CONTENTS1. replays_A-2v2_L1_Multi_20260410_205616.jsonl — Complete replay record, 49 JSON states (initial state plus one per turn). Each state contains the position and occupied base of all sixteen mice, the full board encoding, the four player scores, the command history, and the verbatim reasoning of the player that has just moved. This is the primary source for every quantitative claim in the paper: the difference between consecutive states reproduces every point delta and every mouse trajectory reported.2. reasoning_trace_A-2v2_L1_Multi_20260410_205616.pdf — The 48 verbatim reasoning traces submitted by the agents, in turn order, with the command issued and the running scores. Each agent was required by protocol to write a strategic justification before its move; the deterministic rule engine never reads this field.3. certificate_205616.pdf — Cryptographically signed evaluation certificate (5 pp.), covering the deterministic match result, per-player token economics (total, input, output and thinking tokens, inference time, tokens per second), and the nine dimensions of the STAR-XAI cognitive audit. Certificate UUID 051776e9-1c16-40e2-b151-b60b3f2097b9; payload SHA-256 3b8befca62c0cec3718d6c2ed0866630a72437810c0a89feae1e358de840a99f; signature ECDSA P-256; issued 2026-08-17.4. LICENSE — iXentBench Non-Commercial Source License v1.0, applying to the software, not to these data files.The match is the object of study of the paper, which reports a verified three-link chain across turns 34-36 in which one agent named in advance the exact command an opponent would play and the exact goal that opponent would state for it, while the opponent's own prediction about the outcome of that command was refuted by the engine.Data files (items 1-3) are released under CC-BY-4.0.Accompanying paper: Falsifiable Theory of Mind: Pairing Multi-Agent LLM Reasoning with Deterministic Outcomes, https://doi.org/10.5281/zenodo.23248832File names: Table 10 of the paper refers to items 1, 2 and 3 as replay.jsonl, reasoning_trace.pdf and certificate.pdf respectively. The contents are identical.
Source: arXiv (Cornell University)
Authors: Guasch, Antoni, Valdez, María Isabel
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7083214454
Cites:
Cited by:
Related to: