Answers are grounded in this paper and cite their sources.
Published computational methods routinely stop running within a few years of release. Here we first measure this decay on a principled, seeded random sample of N=20 software tools published in the journal Bioinformatics (2018-2021): running each repository as a naive user would today, only 3 of 20 install and execute, representing an 85% decay rate (95% Wilson CI 64-95%). Second, we introduce Lazarus, an LLM-agent system that resurrects a dead repository from a bare GitHub URL into a callable, independently-verified component. On the same sample, Lazarus revived 17 of 17 dead repositories (100%; 95% Wilson CI 82-100%; 20/20 overall), and 5 of 20 reproduced the paper's own reported metric. Each revival emits a portable contract in the form of an importable module, a CLI, a pinned container, and a smoke test that re-runs on a fresh input. We release a registry of revived tools, upstreamed fixes as two maintainer pull requests, and ship the agent-free measurement as a one-command check (lazarus decay-check). The full run cost a median of $1.16 per repository over ~20.6 hours of wall-clock. In a second, complementary study we apply an agent-free install-time instrument across venues: peer-reviewed research software (JOSS; N=173; cross-domain Python/R) versus software linked from regular journals that do not review code for runnability (N=257). Agent-free, reviewed software packages far more often (95% vs 42%; p<0.001) and, among packages, install-decays less (37% [30-45] vs 61% [51-70]; p=0.0002). Yet when the agent is pointed at the decayed repositories it revives both arms equally: a shipped example ran for ~92% of each and was independently verified for 83% (a purposive one-per-domain reviewed pilot) and 79% (a full unreviewed census). The reviewed/unreviewed gap is therefore packaging discipline, not recoverability. Peer review predicts whether code is installable, not whether the science can be brought back. We foreground the modest samples and the wide, overlapping revival intervals throughout. All metrics regenerate from the released repository https://github.com/DoctorDean/lazarus.
No takes yet. Share an insight, caveat, or question.
Dean Sherry (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: