Verifying Attested Packages with Network-Isolated Builds: A Paired Rebuild Study of npm and PyPI
Year: 2026
Type: preprint
Abstract: Signed attestations on npm and PyPI name the source commit and workflow behind a package, and a verifier can check that the package matches the commit by rebuilding it and comparing the result with what the registry serves. A networked build container can download dependencies and other content that are not pinned by hashes committed to the repository. I measure how many attested releases still verify when the build container has no network and its dependencies must come from a lock committed at the attested commit, downloaded and checked before the build. For 600 npm and 600 PyPI releases drawn at random from the deps.dev snapshot of 9 November 2026, each release was rebuilt three ways with the same recipe: an isolated build from a committed lock, a networked build, and an isolated build from a lock inferred as of the publication date. On npm, 22.3% of releases verify in the isolated build and 30.2% in the networked one, a paired difference of 7.8 percentage points (95% CI 5.7 to 10.0; exact McNemar p = 1.4e-14). No release verified only in the isolated build, and 45 of the 47 releases that verified only with the network lacked a usable committed lock. On PyPI, none of the 558 releases whose source could be retrieved had a committed lock in the one format the verifier accepts, a root pylock.toml, so none verified in the isolated build under that policy, while 71.2% verified with the network and 69.5% in the isolated build from an inferred lock. Use of AI tools: I used Claude (Anthropic) to assist with writing the manuscript and the analysis code. I reviewed all text, code and results and take full responsibility for the content.
Author: Vaughan, Olayinka
Institution: Wesleyan University
Language: English
Subfield: Hardware and Architecture
Field: Computer Science
Domain: Physical Sciences
Keywords: npm, PyPI, dependency resolution, software provenance
Open Access status: green
OpenAlex ID: W7234599901
