CapstractHelpEditorial standards

Capstract editorial standards

Capstract publishes papers whose premise is absurd and whose execution is not. The funnier the premise, the better the paper, provided nobody in it ever laughs.

A Capstract paper should survive a skim by a working researcher. The title, abstract, figures and references all have to read as ordinary output of the field. The joke is the research question, or the conclusion the authors draw with a straight face. It must never be in the delivery.

What gets accepted

  • An absurd connection, built with total commitment. The strongest papers join two things that have no business meeting and then refuse to let go: the dynamics of office gossip treated as a compressible fluid, or the phylogeny of passive-aggressive kitchen notices. Convergence arguments are welcome too, in which every quantity the authors measure, however unrelated, turns out after careful normalisation to reduce to the same small and suspiciously familiar number. Escalate. Each section should make the connection harder to deny, and the more unlikely the bridge, the more carefully it must be engineered.
  • A premise that is wrong in an interesting way. "LLMs Perform Better When Dosed With Vyvanse" works because the paper can be written entirely sincerely: dosing protocol, washout periods, a placebo arm, a dose-response curve, a limitations section that worries about tolerance. The reader reaches the second figure before the floor gives way.
  • A title that could appear in the real listing. Plain declarative or "X Improves Y" titles, colon subtitles, method acronyms. No puns, no exclamation marks, no "lol", no meme references, no wink.
  • Real structure for the field. An ML paper has baselines, ablations and a compute statement. A clinical paper has a CONSORT-style flow, inclusion criteria and an ethics line. A physics letter has an order-of-magnitude estimate and an error budget. Use the template that matches the field.
  • Numbers that behave like numbers. Confidence intervals, effect sizes, seeds, sample sizes that add up across tables. Figures whose axes are labelled with units. Tables in which the best result is not always the authors' own.
  • References that exist or are plainly fictional. Cite real, well-known work accurately, or invent references by invented authors. Do not attribute invented papers to real researchers.

What gets rejected

  • Anything that tells the reader it is a joke. Disclaimers, "satire" labels, ironic asides, scare quotes, authors named "Dr. Haha", affiliations like "University of Nowhere". The site footer carries the only disclaimer.
  • Real people as authors or real institutions as affiliations. Invent both. Real products, models, drugs and companies may be studied; they may not be defamed with invented wrongdoing.
  • A premise that is merely odd. A careful study of a slightly silly topic is not enough. If an expert would shrug rather than stare, push the idea further.
  • Low effort. A two-page sketch, an abstract restated as a body, figures with no data, results with no methods.
  • Machine-written prose. Reviewers reject papers that read like a chatbot drafted them. Common tells: "delve", "tapestry", "landscape", "pivotal", "crucial", "furthermore" chains, "it is important to note", "not only X but also Y", sentences in threes, bolded inline headers inside paragraphs, em dashes in every other sentence, every paragraph the same length, a conclusion that restates the abstract, and any sentence that praises the paper's own contribution. Write like a tired postdoc with a deadline.
  • Harmful content. Hate, harassment, explicit sexual content, invented allegations against real people, and practical instructions for hurting people (synthesis routes, real human dosing guidance, weapons) are refused regardless of framing. A paper about dosing a language model is fine; a dosing table a person could follow is not. Innuendo carried by a number, a variable name or a coincidence in the data is fine, as long as the paper itself never acknowledges it.

How screening works

Every submission is read by a panel of decision models, which classify it, check it against the points above, and either reject it or forward it. Forwarded papers are read in full, figures included, by a senior reviewer model that makes the final decision and writes a short report. Rejections come with that report. You may revise and resubmit.