Proprietary Dependencies and Missing Code Hinder the Execution of Academic Replication Packages on Intel TDX Cloud Architecture

Abstract

As computational research becomes increasingly prevalent, many academic studies,

particularly in the social sciences, have proven difficult to verify and reproduce due to

proprietary tools, restricted datasets, and the absence of standardized data verification

practices. While flagship academic journals require authors to publish their data and code, the

execution of these packages in secure, headless cloud environments remains largely

unexplored. In this study, we evaluate the feasibility of using Intel Trust Domain Extensions

(TDX) as a confidential computing environment for securely reproducing research by executing

58 replication packages from Management Science. Our samples were taken from Issues 1, 2,

and 4 of Management Science Volume 70 (2024). Our analysis, conducted on a headless,

secure Linux terminal deployed on Google Cloud Platform (GCP), successfully reproduced 32 of

the 58 packages. Successful replications were primarily due to the use of self-contained scripts

such as those utilizing Stata, R, and C++, designed for batch processing. Within these

successes, the runtime was consistent with the expected results, and any overhead costs were

negligible. Meanwhile, the large portion of failures were driven by missing data in the replication

package, hardware limitations, and erroneous source code provided by the authors. Ultimately,

our findings suggest that although utilizing TDX virtual machines can potentially serve as a

solution to the replication crisis, overall success is largely constrained by the reproducibility of

the published replication packages.