Modern scientific claims are rarely self-contained. To verify a machine learning model, reproduce a bioinformatics pipeline, or audit a dataset, a reader must step outside the static text of a paper and follow a hyperlink to an external repository. However, the web is fundamentally ephemeral. When scientific literature relies on standard URLs to connect text to its digital bedrock, it inherits a systemic vulnerability: link rot.
The Fragility of Location-Based Knowledge
Why do links rot? Standard URLs are location-based identifiers; they specify where an asset is currently hosted (e.g., a specific folder on a university server or a personal Dropbox account) rather than what the asset is. If a researcher changes institutions, a lab restructures its website, or a commercial cloud service alters its API, the location changes, and the link breaks permanently.
This instability is a pervasive tax on scientific reproducibility. Empirical audits across diverse fields reveal a consistent pattern of decay over time. For instance, in Automatic Analysis of Available Source Code of Top Artificial Intelligence Conference Papers, Jialiang Lin et al. found that approximately 8% of source code links in AI papers become entirely inaccessible or private within just a few years of publication.
In fields with longer-tail software dependencies, the numbers are even more stark. In Challenges and recommendations to improve the installability and archival stability of omics computational tools, Serghei Mangul et al. conducted a systematic review of bioinformatics tools and found that nearly 28% of the URLs published in peer-reviewed papers were completely broken, with older publications suffering the highest rates of failure. When these links fail, the underlying tools lose their archival stability, rendering the science effectively unverifiable. Even high-impact journals are forced to issue formal administrative corrections simply to fix broken or misformatted external links to supplementary materials, as seen in Author Correction: Higher-order topology in bismuth by Frank Schindler et al.
The Expanding Scope of Digital Artifacts
The problem of link rot has intensified as the nature of scientific outputs has evolved. In the early days of the web, the primary challenge was archiving the manuscripts themselves. In The Computing Research Repository: Promoting the Rapid Dissemination and Archiving of Computer Science Research, Joseph Halpern et al. described how the fragmentation of computer science research across decentralized personal homepages and departmental servers threatened the long-term availability of the scientific record—a crisis that prompted the creation of a centralized archive detailed further in CoRR: A Computing Research Repository by Joseph Halpern et al.
Today, the crisis has migrated from the text to the underlying artifacts. Modern research requires off-site hosting for massive datasets and computational models due to severe storage constraints within traditional journal formats, a challenge outlined by Dov Greenbaum et al. in Structuring supplemental materials in support of reproducibility.
In a recent audit of dataset management practices, Yiwei Wu et al. in A Systematic Review of NeurIPS Dataset Management Practices observed a "wild west" of non-persistent hosting solutions, where critical machine learning datasets are frequently distributed via personal lab websites or temporary cloud drives, leading to rampant link rot. Similarly, a meta-analysis by Rui Li et al. in Open-Source Biomedical Image Analysis Models: A Meta-Analysis and Continuous Survey highlighted an "openness gap" in biomedical AI: even when code is shared, the actual trained model weights are often relegated to transient file-sharing platforms, ensuring their rapid decay and permanent loss.
Decoupling Identity from Location
To solve link rot, the scientific ecosystem must transition from location-based addressing (URLs) to registry-based addressing using Persistent Identifiers (PIDs), such as Digital Object Identifiers (DOIs).
The intuition behind a PID is simple: it introduces a layer of indirection. Instead of linking directly to a specific server, a paper cites a permanent string managed by a central registry.
Location-Based (Fragile):
Paper Link ──> [http://university.edu/~lab/data.zip] ──> Server File (Breaks if server changes)
Registry-Based (Persistent):
Paper Link ──> [DOI Registry Entry] ──> [Current Server URL] ──> Server File
▲
(Updated if server changes)
When a reader clicks a DOI, the registry resolves it to the current physical URL of the asset. If the asset moves to a new server, the repository managers update the registry entry. The citation in the paper remains completely unchanged and functional.
The Human Bottleneck and Identifier Leakage
However, building persistent infrastructure is only half the battle; human behavior frequently undermines it. Even when stable repositories and PIDs are available, the ecosystem suffers from what Henry S. Thompson terms "identifier leakage" in Can Common Crawl reliably track persistent identifier (PID) use over time?.
Identifier leakage occurs when a user navigates to a persistent identifier, allows their browser to resolve it to the final destination URL, and then copies that destination URL from the address bar to share or cite in subsequent work.
User clicks DOI ──> Resolves to Destination URL ──> User copies Destination URL ──> New Fragile Citation
By substituting the permanent link with an unstable, direct URL, users inadvertently re-introduce fragility into the web ecosystem. The author notes that this behavioral pattern causes stable PIDs to lose their efficacy over time.
The scale of this gap is visible across the broader web. In an analysis of English Wikipedia's scholarly foundations, Harshdeep Singh et al. established in Wikipedia citations: A comprehensive data set of citations with identifiers extracted from English Wikipedia that a massive persistent identifier gap exists: the vast majority of external citations on the platform rely on raw, direct URLs rather than stable markers like DOIs or ISBNs. This systemic reliance on transient links leaves public knowledge bases highly vulnerable to link decay, complicating the long-term verifiability of information.
True reproducibility requires treating code, data, and models not as ephemeral afterthoughts, but as first-class citizens of the permanent scientific record, bound by persistent infrastructure and disciplined citation practices.
Key Foundational Work
The Computing Research Repository: Promoting the Rapid Dissemination and Archiving of Computer Science Research (1998) — contributed 1 concept elements to this direction.
CoRR: A Computing Research Repository (2000) — contributed 1 concept elements to this direction.