Paper Fabric · October 2026

Closed Source: The God that Failed

Reverse engineering, espionage and the ethics of citation in the age of coding agents

6 works, 5 DOIs.

Click a tile to open its DOI, or the work's own page when it has no DOI yet. A DOI still resolves after a repository moves, so cite the DOI, not the URL.

  • non-profitideas

    closed-source

    closed source: the god that failed: the paper you are reading about. its latest DOI always opens the newest release.

    DOI 10.5281/zenodo.23333656

    read the paper (PDF) · permanent DOI for v0.3.0: 10.5281/zenodo.23349831

  • non-profitwater

    marola

    an open, non-profit platform for sea conditions and bathing-water quality at Brazilian beaches

    DOI 10.5281/zenodo.23224155

    repository

  • non-profitwaves

    ww3-gpu

    the WW3 GPU Lab: moving WAVEWATCH III kernels to GPUs without changing the answer

    DOI 10.5281/zenodo.23221351

    repository

  • profitnix

    picasso

    self-contained Nix flakes for scientific and agentic toolchains

    DOI 10.5281/zenodo.23311208

    repository

  • bookgame theory

    nashkell

    an open book in progress, in the voice of Nash: game theory, pictures that are equations, and the mask of Latin that Section 5 studies

    DOI 10.5281/zenodo.23356739

    repository

  • profit (?)ideasfriends

    ouroboric

    open factories that improve themselves, by Bruno Valério. a company whose charter says that making money is not a goal for now. the ouroboric ideas in this paper are his.

    DOI none yet

    github.com/ouroboric-dev

What changed in v0.3.0

every release

The latest release, v0.3.0 on GitHub, with its permanent DOI 10.5281/zenodo.23349831.

  • on Zenodo the record is a book, not a preprint (publication_type in .zenodo.json); the generated citations follow: CFF type: book, BibTeX @book, APA "[Book]"
  • the website publishes the paper as paper.pdf, a name that does not change, and shows a preview of it in "Read"; the Zenodo record and the READMEs link to that preview
  • the record's description names the writer's three other works with their DOIs (marola, WW3 GPU Lab, picasso) and says whose the Ouroboric ideas are
  • the record's ties add requires picasso and cites nashkell (#44), and references Hoppe's Democracy: The God That Failed by its publisher's DOI
  • the record's description starts with the book as a PDF: the link to read it and the link to download it
  • three analysers check the Markdown files (just md-quality), and a badge in the README says if every link is alive; two more count the links to GitHub and to Zenodo
  • the website has a CI card: what happens when the book changes, with the state of each check
  • the 50 thinkers are in a new order, with no two neighbours of the same area
  • the repository has a social preview: the website's cover as a picture for a link in a chat

The 50 Thinkers

50THINKERS.md

The book is 100 pages: 50 thinkers, two first-person pages each. 43 are proposed, 7 places are still open.

  1. 1Mário Ferreira dos SantosBrazilian philosopher, *Filosofia Concreta* (1957); Pythagorean (*Pitágoras e o Tema do Número*, 1956)
  2. 2Austin Kleon*Steal Like an Artist* (2012; *Roube como um Artista*, Rocco, 2013): creative work builds on the work of others
  3. 3Socratesmathematics
  4. 4Gregory Wolfe*Beauty Will Save the World* (2011; *A Beleza Salvará o Mundo*): recovering the human in an ideological age
  5. 5Edmund Burkelater named the founder of conservatism; in 1775 he opposed the extension of Watt's steam-engine patent, and backed the extension of a porcelain patent (notes/202610130526)
  6. 6Linus TorvaldsLinux and Git: openness that scales (Section 2)
  7. 7Open
  8. 8Boaz Barakno general obfuscator can exist (Section 5a)
  9. 9Julia Cameron*The Artist's Way* (1992; *O Caminho do Artista*): creativity as a daily practice
  10. 10Dante Alighierireason to come
  11. 11Nicholas Carlinistealing part of a production language model (Section 4)
  12. 12Open
  13. 13Hans-Hermann Hoppethe title formula; ideas cannot be property (BOOKS.md)
  14. 14Marc Stevensthe first SHA-1 collision, why Git moved on
  15. 15Florian Tramèrmodel extraction through prediction APIs (Section 4)
  16. 16Michele BoldrinAgainst Intellectual Monopoly (BOOKS.md)
  17. 17Arfon M. Smithsoftware citation principles (Section 6)
  18. 18Olavo de Carvalho*O Jardim das Aflições* (1995), and *Símbolos e Mitos no Filme "O Silêncio dos Inocentes"*
  19. 19David K. LevineAgainst Intellectual Monopoly (BOOKS.md)
  20. 20John F. Nashsecurity from computational cost, the 1955 letters (Section 5a)
  21. 21Geoffrey Hintonknowledge distillation (Section 4)
  22. 22Stephan KinsellaAgainst Intellectual Property (BOOKS.md)
  23. 23Ronald L. Rivestpublic-key cryptography: secrecy only of the key
  24. 24Open
  25. 25Luiz Felipe PondéBrazilian philosopher; reason to come
  26. 26Auguste Kerckhoffsa system must stay secure when everything but the key is public
  27. 27Open
  28. 28Ciceromathematics
  29. 29Open
  30. 30Robert M. PirsigQuality, and maintaining your own machine (BOOKS.md)
  31. 31Gwern Branwenpublished the Nash–NSA letters online (Section 5a)
  32. 32Open
  33. 33Jiang Xueqinthe Predictive History lectures (YouTube, since 2023); reason to come
  34. 34Hideaki Annocreator of *Neon Genesis Evangelion*, the most classical anime in history
  35. 35Sandy Maguirethe isovector book build behind Paper Fabric (Section 7)
  36. 36Adi Shamirpublic-key cryptography: secrecy only of the key
  37. 37Theodore Dalrympleto be written later
  38. 38Mark D. Wilkinsonthe FAIR principles (Section 6)
  39. 39Stanley Kubrick"the peer reviewer destroyer" (the owner): *Eyes Wide Shut* (1999), his last film, a work on secrets (`notes/202610140034-peer-reviewing-the-god-that-failed.md`)
  40. 40Isaac Asimovreason to come
  41. 41personXwho shares, multiplies
  42. 42Platomathematics
  43. 43Pamela Samuelsonlaw and economics of reverse engineering
  44. 44Eric S. Raymondthe cathedral and the bazaar (Section 2, related work)
  45. 45Open
  46. 46Marcus Baffato be explained
  47. 47George Orwellreason to come
  48. 48Aristotlemathematics
  49. 49Suzanne Scotchmerlaw and economics of reverse engineering
  50. 50Michel de Montaignethe *Essays*: why one writes at all

Second brain

notes/index.md

A Zettelkasten that Obsidian opens: one idea per note, linked, each with its verification status. The book is written from it. A note marked approved is one the owner approved for the book.

32notes
3essays

v0 (preprint, not peer reviewed)

  • The paper (PDF)

    Typeset with XeLaTeX from the Markdown sources.

  • ASD-STE100 text (PDF)

    The same paper in Simplified Technical English, with the same citations and numbers.

  • Edição em português (PDF)

    The printed edition in Brazilian Portuguese, with the same citations and numbers.

  • Releases

    Each tagged release: the PDFs and the arXiv source bundle. Zenodo archives each one and gives it a permanent DOI.

The preview does not show on every phone. Open the PDF or download the PDF.

14pages
11sections
41references
34references fetched and verified
1author

Abstract

For thirty years, closed source has promised that secrecy protects value. I argue that the promise has failed, for science first. Three trends from 2024 to 2026 make the case. Large language models now decompile binaries into re-executable source. Black-box queries extract parts of production models, and distillation copies capabilities at industrial scale. The largest 2026 trade-secret dispute in artificial intelligence, Apple’s suit against OpenAI, turns on people, laptops and recruiting, not on code that stayed secret. Against this background I revisit Linus Torvalds’ Linux and Git as evidence that openness scales, and my closed undergraduate thesis as evidence of what secrecy costs a researcher. I review obfuscation, including my own, and why it cannot hide code, and I contrast it with John Nash’s 1955 idea of security from computational cost: signed commits, keys held in hardware and two-factor authentication protect what secrecy cannot. I then argue that the ethical answer to universal reverse engineering is professional citation: persistent identifiers, machine-readable citation metadata and agents that borrow, cite and propagate credit by default. I describe Paper Fabric, an open, self-improving pipeline in the Ouroboric style that produced this preprint, and list its own known limitations.

  • open source
  • reverse engineering
  • AI-assisted reverse engineering
  • model distillation
  • trade secrets
  • software citation
  • research software
  • FAIR
  • coding agents

Contents

  1. 1

    Introduction

  2. 2

    The bazaar that worked: Linus Torvalds, Linux and Git

  3. 3

    A closed-source confession: Chemo-RX

  4. 4

    Science and espionage in 2026: the reverse-engineering explosion

  5. 5

    Why closed source is dead: five lines of evidence

  6. 6

    Obfuscation, Nash's letters and how to secure source code in 2026

  7. 7

    The ethical answer: professional citation

  8. 8

    Paper Fabric: an ouroboric factory for papers

  9. 9

    Related work by the same author

  10. 10

    Limitations

  11. 11

    Conclusion

Author

How it is made

Paper Fabric

Paper Fabric is the open pipeline that produced this preprint. Humans steer and agents build: authors set the thesis and write their own accounts, agents research, draft and check, and every source is fetched before it is cited.

  1. Write. Markdown prose, one file per section, and a BibTeX file.
  2. Build. pandoc with pandoc-crossref and citeproc, a LaTeX template, XeLaTeX. An unresolved citation stops the build.
  3. Check. Every reference has a row in the reference log; the ASD-STE100 text keeps every citation and number.
  4. Release. A tag makes a GitHub release with the PDFs and the arXiv source. Zenodo archives the release and mints its permanent DOI a few minutes later.

Tools pinned by Nix from picasso's labs: paper for the PDF and gleam-website for this site. The build follows Sandy Maguire's books (isovector).

every changestate now

what happens when the book changes

a change to the book is a proposal first. machines read it, then a person decides. each badge is the state of that check now, drawn from GitHub's own data.

  1. a change is proposed

    a pull request on GitHub: the text that changes, and why.

  2. five checks read it

    each one reads the change before the merge. a red one is a reason not to merge.

    • pdf: the state of its last runbuilds the three PDFs. a citation with no source stops it. checks every reference, the simplified english and the portuguese.
    • site: the state of its last runbuilds this website from the book's own files and runs its tests.
    • citation: the state of its last runchecks the record that Zenodo and GitHub read, so the work stays citable.
    • .md: the state of its last runthree analysers read the markdown: structure, plain words, links to files.
    • live links: the state of its last runevery web link of the markdown and of this website is fetched. green means 100% answer.
  3. a person merges, and it is published

    the rule here: no agent merges. after the merge:

    • site: the state of its last runafter the merge, the website is published again, with the new PDF.
    • live links: the state of its last runthe links are fetched again, and each Monday: a page on the web can go away with no change here.
    • release: the state of its last runwhen the owner tags a version: the PDFs go to a GitHub release, and Zenodo mints a permanent DOI.

all runs on GitHub · what the checks do not check

Citation scores

CITATION_SCORES.md

How much of each README is cited: the share of its sentences that carry a link, a DOI or another citation mark. A README is not a paper, and a low score is not a fault; it says how much of a text points at something else.

  • closed-source h0ffmann/closed-source README.md, working tree

    30% cited 35 of 115 sentences. THIN. Most of it is yours, or unsupported.

  • marola marola-dev/marola README.md, cf7bd4f

    48% cited 57 of 120 sentences. THIN. Most of it is yours, or unsupported.

  • ouroboric-dev ouroboric-dev/.github profile/README.md, b799793

    50% cited 2 of 4 sentences. FAIR. Half of it stands on a source; the rest is yours, or unsupported.

  • h0ffmann h0ffmann/h0ffmann README.md, 5d81ea1

    73% cited 22 of 30 sentences. FAIR. Half of it stands on a source; the rest is yours, or unsupported.

  • picasso h0ffmann/picasso README.md, 29eb385

    42% cited 60 of 142 sentences. THIN. Most of it is yours, or unsupported.

Measured with scripts/citation-score.sh, which counts marks and cannot tell if a citation is right. Each row names the commit that was scored.

Cite

Professional citation is the paper's own argument. Cite it, and cite what you borrow.

Latest DOI10.5281/zenodo.23333656
One DOI for the whole work. It always resolves to the newest release. Cite this one.
Permanent DOI · v0.3.010.5281/zenodo.23349831
One DOI for release v0.3.0 alone: the same text forever. Add it when the exact wording matters, as in a quotation or a review.

Each release gets its own permanent DOI: v0.3.0 10.5281/zenodo.23349831 · v0.2.0 10.5281/zenodo.23335204 · v0.1.0 10.5281/zenodo.23333657 · all versions on Zenodo

@book{hoffmann_2026_closedsource,
  author    = {Hoffmann, Matheus},
  title     = {{Closed Source: The God that Failed}},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.23333656},
  url       = {https://github.com/h0ffmann/closed-source}
}