The frozen record
The scientific paper has not changed in form since the first issue of Philosophical Transactions in 1665. It is text with figures. For three and a half centuries, this was the only practical option: print is static, and static is what could be reproduced and distributed at scale.
That constraint no longer exists. The web can render arbitrary computation. A figure can respond. A model can be explored. An analysis can be reproduced in the reader's browser, with the reader's own parameters, without installing anything. And yet the default scientific artifact in 2026 remains a PDF: a simulation of paper, shipped as a file.
PDFs are excellent for plenty of things: archival records, portable documents, anything designed to be printed, signed, and filed. A letter, a legal brief, an annual report are all well served by a fixed, reproducible format. Where that format struggles is with arguments that depend on showing how a system behaves across a range of conditions, or with intuition that can only be built by handling something directly and watching what changes.
What static figures conceal
Scientists working with high-dimensional data routinely build and explore three-dimensional representations (dimensionality reductions, spatial cell maps, feature spaces) as part of the analysis itself. Rotating the point cloud is part of how the structure is understood. You spin it until the groups separate cleanly, until the elongation axis becomes legible, until you understand where the boundary is and what it looks like from the side.
Publishing collapses this into two or three frozen panels chosen by the author at the time of writing. The reader receives those angles and nothing else. They cannot rotate to a direction relevant to their own question. The author's navigation of the 3D structure has been compressed into a fixed sequence of snapshots; the reader is left to reconstruct the rest from what was chosen for them.
Figure 1. Two experimental conditions in a three-component feature space. A published paper would present this data as two or three fixed-angle panels chosen by the author.
The same geometric problem applies when the data itself lives on a three-dimensional surface. Cell migration direction is a unit vector in ℝ³, which is a point on S². The analysis of how this distribution changes over time is done by rotating the sphere, exactly as above. A published paper compresses this to one or two rose plots from fixed projections, discarding the elevation component entirely. The time course is reduced to a before-and-after pair at best. What is lost in that compression is often the most informative part of the measurement.
Figure 2. Directional distribution of 3D cell migration, represented as a density on the unit sphere. Each point on the surface corresponds to a possible migration direction; colour encodes the fraction of cells moving in that direction at the selected time. The distribution begins near-uniform and concentrates toward the chemoattractant source as stimulation time increases.
The sphere is the actual geometric home of this data, not a metaphor for it. A 2D projection of directional data on S² always discards one dimension. For a symmetric, unimodal distribution the loss may be acceptable. For an asymmetric or multimodal distribution, which is common when cells are navigating competing gradients or structured matrix, it is not. The interactive version recovers what the projection discards.
Exploration as comprehension
A different kind of loss occurs with mathematical or computational models. A static figure shows one point in a parameter space. The reader is left to imagine the rest.
Consider a simple signal detection scenario: two populations (signal and noise) modelled as Gaussians with different means. A detector flags observations above a threshold. The relevant quantities (sensitivity, specificity, the tradeoff between them) depend on where the threshold is set and how much the distributions overlap. This is the foundation of receiver operating characteristic (ROC) analysis, used in clinical diagnostics, machine learning evaluation, and communication systems.
A textbook figure shows one configuration. The figure below makes the full parameter space accessible. The intuition that builds from direct manipulation of these two variables is genuinely difficult to acquire from any static figure, however carefully annotated.
Figure 3. Left: two Gaussian distributions (noise in blue, signal in green) separated by an adjustable mean distance, with an adjustable detection threshold. Right: the corresponding ROC curve, tracing every possible operating point as the threshold varies. The green dot marks the current setting.
A reader who has moved both sliders will interpret ROC curves in future papers differently than one who has only read a definition. They have watched the operating point trace the curve as the threshold shifts, and seen AUC change as the distributions pull apart. That embodied sense of what a given AUC value implies about the underlying overlap is harder to acquire from a static figure and a formula.
Geometry that requires a viewpoint
A fixed camera angle is a choice about what to conceal. When the object of interest is three-dimensional, a single projection will always hide some of its structure, and the relevant view is rarely known in advance. Static figures can show multiple panels from chosen angles, but they cannot show all angles, and they cannot let the reader choose.
Consider a likelihood surface: the log-likelihood of a dataset as a function of two model parameters. The topology of this surface determines everything about how estimation behaves. A smooth unimodal bowl converges reliably from any starting point. A surface with ridges, saddle points, and competing valleys does not. Which regime a given model operates in is not readable from the estimated parameters alone; it requires seeing the surface.
The figure below shows such a surface. The time slider traces an optimisation trajectory across it.
Figure 4. A log-likelihood surface over two parameters, with a global minimum near (−1.6, −1.6) and a shallower local minimum at (1.469, 1.469). The trajectory slider traces a gradient descent path across the surface. Start A becomes trapped on the ridge separating the two basins; Start B descends into the local minimum at (1.469, 1.469). Neither trajectory reaches the global minimum.
Two things are only visible in three dimensions. The ridge separating the two valleys is substantially steeper on one side than the other; from directly above it would appear nearly symmetric. Start A stalls on this ridge, unable to commit to either basin. Start B clears the ridge but descends into the shallower valley, settling at the local minimum (1.469, 1.469).
The fourth dimension, time, turns a static surface into a record of a process. A static paper can mark the start point, the endpoint, and perhaps a few intermediate positions indicated by arrows. The full path through a three-dimensional landscape requires a medium that is not paper.
When the format is the argument
The four figures above carry the argument rather than illustrating it. Take them away and the article loses its evidence, not merely its examples; what made the claim demonstrable disappears with them. The interactivity is doing epistemic work here.
Vivum evaluates submissions on a single question: would replacing the interactive elements with static screenshots lose something? If yes, the format is doing real work.
The tradition this builds on is not new. Bret Victor's essays have been making this argument since 2011. 1 Victor's "Explorable Explanations" (2011) introduced the term and the concept. His subsequent essays, particularly "Up and Down the Ladder of Abstraction" (2011) and "Learnable Programming" (2012), developed the idea that the ability to vary parameters and see consequences is a fundamentally different epistemic tool from reading a description of the same system. Distill published machine learning research in this format from 2016 to 2021 and produced some of the most-read papers in the field. 2 Distill's "Feature Visualization" (Olah et al., 2017) and "The Building Blocks of Interpretability" (Olah et al., 2018) reached audiences far outside the ML research community, in part because their interactive figures allowed non-specialists to explore the results directly. Distill ceased accepting new submissions in 2021, citing the maintenance burden of the format. The gap it left is part of what Vivum is designed to address. The approach has been proven repeatedly. What has never existed is a permanent, multi-field home for it.
Vivum fills that gap: an open index for web-native articles across any scientific field, neither a journal nor a host, but a persistent and discoverable place where this kind of work can be found, discussed, and formally reviewed when the community decides it warrants it.
How to write one
The tooling has improved considerably. You do not need to build a custom JavaScript framework. The most common paths are:
Distill template: Apache 2.0 licensed HTML framework with built-in support for sidenotes, citations, responsive figures, and KaTeX. The standard starting point for ML and interpretability work.
Observable notebooks: reactive JavaScript in the browser, ideal for data exploration and statistical visualisation. Notebooks can be published directly as standalone URLs.
Quarto / R Markdown / Shiny: if your analysis lives in R, these tools render interactive HTML documents with minimal additional work. Shiny apps can be hosted for free on shinyapps.io.
Plotly Dash: Python-native framework well suited to biomedical data and genomics. Produces deployable web applications with minimal boilerplate.
Jupyter Book: for Python-based analysis, Jupyter Book produces a complete static site from notebook files, with optional interactive widgets via ipywidgets.
Plain HTML and JavaScript: what this article is built with. D3, Vega-Lite, or Plotly can be dropped in via CDN. No build step.
Vivum HTML template: a ready-to-use single-file template styled to match Vivum. Includes support for KaTeX maths, D3 figures, interactive controls, sidenotes, a sticky TOC, and a citation block. Host it anywhere and submit the URL. Available on the How to publish page.
The only requirement Vivum imposes is that the article lives at a public URL and that the interactive format does genuine work. Authors host wherever they like (GitHub Pages, Netlify, Vercel) and submit the link. They retain full ownership. Vivum stores the metadata, hosts the discussion thread, and connects the article to the review process when the community judges it ready.
References
- 1. Victor, B. (2011). Explorable Explanations. worrydream.com/ExplorableExplanations
- 2. Olah, C., Mordvintsev, A., & Schubert, L. (2017). Feature Visualization. Distill. doi:10.23915/distill.00007
- 3. Distill editorial team. (2021). Distill hiatus. distill.pub/2021/distill-hiatus
Cite this article
Boix Campos, J. (2026). The Living Article. Vivum. https://vivum-pub.org/editorial/the-living-article
@article{boixcampos2026living,
author = {Boix Campos, Javier},
title = {The Living Article},
journal = {Vivum},
year = {2026},
url = {https://vivum-pub.org/editorial/the-living-article},
}