Same Trajectories, Different Rankings — audit records and code for a protocol audit of latent world-model planning on the LeWM benchmark
Year: 2026
Type: dataset
Abstract: Audit records, preregistrations, case manifests, frozen environments, workers, patches, and CPU analysis code for public LeWM, Fast-LeWM, AP, and AD-WM checkpoints on the four-task LeWM benchmark, using stable-worldmodel commit 63988116. The archive includes original and corrected continuous Reacher trajectories: the correction disables success termination in this study's continuation adapter, preventing an unintended internal reset after success; released first-arrival evaluation is unaffected. Derived records are CC BY 4.0 and the author's audit code is MIT, with third-party notices retained. Code mirror: https://github.com/2GyounnnG/lewm-protocol-audit. Public weights and input datasets are not redistributed; source URLs and SHA-256 identify them. Corrected endpoint and sustained-window comparisons are descriptive.Version 2 adds Revision 7 A/B/C sensitivity, code audit, native AP renderer replication, scientific logs, trajectories and saved-image pixel comparisons. The published base files are retained. No manuscript, input datasets, model weights or credentials are distributed.
Author: Wang, Ruiqi
Institution: University of California San Diego
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7221699252