Science is built on a simple expectation. If a result is real, others should be able to reproduce it.


Yet across multiple disciplines, that expectation has become increasingly difficult to guarantee. Many published findings cannot be independently replicated, even when peer review has been completed.


The scale of the problem is difficult to ignore. A survey of 1,576 researchers published in Nature found that more than 70% of scientists had tried and failed to reproduce another researcher’s experiments, and over half reported failing to reproduce their own work.¹ Long before the phrase “reproducibility crisis” entered mainstream scientific discussion, John Ioannidis argued that many published findings are vulnerable to distortion when studies are underpowered, analyses are flexible, and publication systems reward positive results over verification.²


His broader point was that scientific systems reward novelty more readily than verification, and those incentives remain deeply embedded.


Survey data suggests many researchers still see reproducibility and replicability as constrained not only by incentives, but also by practical limitations such as time, funding, and access to resources.³


A recent study from January 2026 concluded that the reproducibility crisis in medical education stems from methodological flaws, bias, and systemic pressures, and requires improvements in research methods, academic culture, and oversight to restore scientific credibility.⁴


Today many papers include code repositories, supplementary methods, and data availability statements. In theory, this should make independent verification easier. In practice, true reproducibility often remains difficult. Computational pipelines depend on software versions, hidden preprocessing steps, unavailable hardware, proprietary dependencies, and execution environments that are rarely captured in full. A reviewer may be able to inspect code without being able to run it meaningfully.


Increasingly, reproducibility is constrained not only by transparency, but by access to compute. This matters because modern science has become computationally intensive. In machine learning, neuroimaging, genomics, proteomics, and multimodal biomedical research, many analyses now depend on GPU based workflows that are difficult to recreate outside the originating laboratory. Peer review often remains conceptual rather than executable because the technical burden of rerunning complex analyses is simply too high.


The result is that trust often substitutes for direct validation.


This is where scientific infrastructure begins to matter more than critique alone.


Decentralized science, often described as open science 2.0, may have a more important contribution to make. Its strongest potential lies in building infrastructure that reduces friction between scientific claims and scientific verification.


At AxonDAO, this idea is taking shape through AxonHub, an infrastructure layer designed to make research executable rather than simply readable. The premise is straightforward: a scientific paper should not remain a static document detached from the computational process that generated its findings. Instead, manuscripts can be linked to runnable code, reproducible environments, and accessible compute resources so that results become testable rather than assumed.


AxonHub sits at the interface between scientific publication and computational execution. Researchers submit code alongside their work, while underlying compute is supported through AxonGPU, a GPU layer designed to make high performance scientific workflows more accessible. This means that a reviewer or independent scientist could move beyond reading methods sections and directly test whether submitted code reproduces reported figures, statistics, or model outputs in the same environment in which they were originally generated.


The goal is not simply openness. It is executable trust.


This distinction matters because open code alone is often insufficient. A public repository without accessible compute still leaves many analyses effectively untestable, particularly when workflows rely on large models, GPU acceleration, or complex software stacks. As scientific methods become increasingly AI native, access to reproducible compute may become as important as access to the manuscript itself.


The implications extend beyond peer review. Scientific claims could remain continuously testable after publication. Independent groups could rerun analyses, compare alternative pipelines, and challenge assumptions without rebuilding environments from scratch. Negative findings and failed replications could remain attached to the same scientific object rather than disappearing into unpublished silence.


The infrastructure begins to support science as a living process rather than a frozen publication. Traditional academia still provides what decentralized systems cannot replace - ethical oversight, institutional continuity, domain expertise, and public trust. But institutional science has not yet solved how to make computational verification frictionless at scale.


As science becomes more model driven, multimodal, and computationally intensive, reproducibility will depend less on whether code is nominally available and more on whether scientific communities can realistically execute and interrogate that code.


The reproducibility crisis is often framed as a problem of incentives, publishing norms, and academic pressure but it may also be an engineering problem. Science has evolved into a computational discipline. Its methods of verification must now reflect that transformation. If that happens, the next major contribution from decentralized science may not simply be new funding models or tokenized communities. It may be the creation of infrastructure where scientific claims become easier to run, inspect, challenge, and ultimately trust.


References


1. Baker, M. 1,500 scientists lift the lid on reproducibility. *Nature* 533, 452–454 (2016).

2. Ioannidis, J.P.A. Why most published research findings are false. *PLoS Med.* 2, e124 (2005).

3. Chakravorti, T., Koneru, S., Rajtmajer, S. Reproducibility and replicability in research: What 452 professors think in Universities across the USA and India. *PLOS ONE* 20(3): e0319334 (2025).

4. Ahmady, S., Kohan, N., Hamidi, H. et al. Interpretations of reproducibility crisis in medical education research: a qualitative study. *Sci Rep* 16, 4489 (2026).