Persistent Recursive Worlds Enable Autonomous Software Evolution
Huang, Liang, Zheng and Cheng · arXiv 2608.10450
The outcome is the verdict on the one claim below, under the protocol we fixed in advance. It is not a verdict on the paper as a whole.
Claim tested
Rust driver is faster than the Fortran driver on six workloads (bounds 1.1/4.5/2.5/1.0/1.1/1.0).
What we ran and what came back
The paper reports that its system's Rust rewrite of 13 modules of the MESA stellar-evolution code runs 1.55 to 6.87 times faster than the Fortran original on six numerical workloads. We built both at a pinned commit and timed 75 alternating pairs per workload on one pinned cloud CPU core. Rust was faster on all six, and each ratio cleared its bound, which we set at 0.7 of the lower of the authors' two published figures because those two disagree by up to 1.5 times. The sizes are smaller than the paper's table on five of six workloads: interpolation 1.16 (paper 1.58), ROS2 solver 3.53 (5.30), reaction network 1.15 (1.55), equation of state 1.43 (1.60), opacity 1.35 (1.98). The Newton solver is close at 6.34 (6.87). So Replicated here means at least the stated lower bound on every workload, not the paper's numbers. We used a different Fortran compiler version and reference linear-algebra libraries in place of the ones the authors used, which could move the Fortran side. One machine, one build.
What this result supports
The Rust port's per-workload speedups hold on a shared cloud CPU.
What it does not support
Other hardware; full-simulation wall time.
Record
Protocol, deposited before the run: 10.5281/zenodo.23543344
The authors were sent this result. They have not reviewed this listing.
How the protocol works is on the method page. If you are an author of this paper and something here is wrong, write to customers@aegismind.app and we will correct it on this page.