When the work is the output, not the paper

27 June 26

researchai

Earlier this month I gave my livecoding gigs DOIs. That was the easy case, in one narrow sense: a livecoding gig has no paper. The performance is the whole of the output, so there was never much question about what the citable thing was. The rest of the Cybernetic Studio’s work is messier. PANIC! has two papers attached to it; LLMs Unplugged has one. Once a paper exists, the temptation is to let the paper be the research output and to file the installation as the thing the paper is about.

For these works that gets it the wrong way round. The work itself is the output, whatever physical form it takes, and the paper only documents it.

So this is the second wave: turning the Cybernetic Studio’s creative-AI works into citable research outputs the way I did the gigs. There are seven of them. PANIC! wires generative models into a feedback loop and lets visitors watch the meaning drift; the Perceptron Apparatus is a hand-cranked machine that performs a small neural network’s arithmetic by hand; LLM Brainscan paints every weight of a transformer onto an 8K display as it trains. Slop Salon is a collective of AI agents making art in public; the APS AI Transparency Tracker scrapes every Commonwealth agency’s AI statement daily and commits the diffs to git; Imaginative Restoration was built with NIDA and the National Film and Sound Archive; and LLMs Unplugged is a shelf of literary language models printed and bound as books. Each now has a concept DOI through Zenodo, the all-versions identifier that redirects to the latest archived release, so a citation doesn’t go stale as a living work keeps shipping.¹

Most of these works live in GitHub repos, and Zenodo has a tidy integration for that: tag a release, and it archives a source-zip and mints a fresh version DOI on its own. But this is a pretty ‘software’ workflow: every automatic deposit comes out typed as Software and licensed MIT. For a repo of code that is exactly right, because a source-zip of the tracker honestly is software and honestly is MIT. For an installation that’s a bit of a lie. PANIC! isn’t software. It is an installation that happens to have software inside it, and its archive shouldn’t introduce itself as a code library.

The fix is a .zenodo.json committed to each repo, setting the type, licence, creators, keywords and related identifiers that every future release will carry. The Zenodo-GitHub integration doesn’t do this: it reads .zenodo.json and ignores CITATION.cff. Several of my repos had .cff files sitting in the root, looking authoritative and controlling precisely nothing about the deposit.²

The licence thing is subtle, because once you’ve decided a thing is “creative”, the instinct is to reach for a Creative Commons licence, and for art the reflex is CC-BY-NC-SA: no commercial use, share alike. You must never put a non-commercial licence on code. The NC clause contradicts the open-source grant the code already carries. MIT says do anything with the code, including sell it; NC says you may not; one set of files cannot say both. So the rule I settled on is to let the licence follow the bytes. If the deposit is a source-zip it is Software and it is MIT, full stop, however creative the work it stands for. CC-BY-NC-SA is reserved for the deposits that are themselves creative or teaching artefacts: the LLMs Unplugged books through Cybernetic Studio Press, not the code that typesets them.

The companion papers get linked rather than cited. PANIC! carries related identifiers pointing at two things: its NIME 2025 proceedings entry, and the IEEE SMC 2025 paper that runs a persistent-homology analysis over the same installation runs. Both are tagged isDocumentedBy. LLMs Unplugged points at its ACE 2026 paper the same way. The installation record points at the papers that document it (rather than a paper record laying claim to the artwork).

A few things don’t come out clean. The linking only runs one way under my control. I can point my own Zenodo records at the NIME, IEEE and ACM DOIs, but I can’t make theirs point back, because those records belong to the publishers and the conference communities and I’m a guest in them. The cross-reference is half a handshake. The credit is lossy, too. The point of listing creators in .zenodo.json is to attach ORCIDs, so authorship is machine-readable and unambiguous, but most of my collaborators on these works are practitioners and fabricators who simply aren’t in the ORCID registry. A search this June turned up exactly one claimable ORCID among them. For everyone else the record carries a name and nothing more.

The most awkward gap is that Zenodo has no upload type for “interactive installation”. The list runs to software, dataset, publication, image, video and a catch-all “other”, and that’s the lot. An 8K data-art piece and a room-sized participatory installation both have to file as either software or “other”. It’s software when a source-zip is the honest deposit, and “other” when nothing zippable really stands in for the work. The closest honest mapping wins, but none of them is the true thing. This is the same shape of problem the gigs had. The umbrella I wanted to call a “collection” had to be typed “other”, because Zenodo has no such category. The vocabulary runs out before the work does.

None of this teaches Zenodo what an installation is. What the .zenodo.json files gives me is records that describe themselves accurately: code filed as code, creative work filed as creative work. The seven works are gathered into a Cybernetic Studio community alongside the livecoding one, so the whole portfolio sits in a single place. When the metadata is still wrong, and in a couple of places it is, it’s wrong because Zenodo ran out of words, not because I couldn’t be bothered getting it right.

@online{swift2026whenTheWorkIsTheOutputNotThePaper,
  author = {Ben Swift},
  title = {When the work is the output, not the paper},
  url = {https://benswift.me/blog/2026/06/27/when-the-work-is-the-output-not-the-paper/},
  year = {2026},
  month = {06},
  note = {AT-URI: at://did:plc:tevykrhi4kibtsipzci76d76/site.standard.document/2026-06-27-when-the-work-is-the-output-not-the-paper},
}