Skip to content

Reference pattern: Eduard Cristea's dojo.md

Written 2026-09-13. The publishing shape of this repo was modelled on github.com/edholofy/dojo.md (Eduard Cristea, Holofy, MIT, first commit 2026-02-27, ~10 stars at the time of writing). This note records what we copied, what that repo does not do, and what we added on top.

What dojo.md does

Root layout, as fetched on 2026-09-13:

README.md          11.6 KB   product-style README (see below)
paper.md           32.5 KB   a full paper, at the repo root
CLAUDE.md          11.4 KB   agent-facing project guide (commands, architecture, key files)
LICENSE            MIT
hero.jpeg          hero image referenced from the README
.github/workflows/ci.yml    build + test on Node 18/20/22
courses/  src/  website/    the product itself, 7,400 files

README shape. Hero image at the top. One-sentence problem statement in bold ("Your agent demos well. It fails in production."). A transcript-style demo block showing five iterations of the loop and the score climbing. "What just happened" narrative. A "paste this into Claude Code" quick start with zero setup. A cost table. A leaderboard block rendered as ASCII bars. A table of everything the product covers. A format spec (the SKILL.md standard). CLI reference tables. Development commands. A three-line "Mission" (today, tomorrow, future). License.

paper.md shape. Title, author, email, month. Abstract. 1 Introduction (problem framed as two gaps; four core ideas as a numbered list). 2 Related Work (seven short subsections, each ending with how dojo.md differs). 3 System Architecture (pipeline diagram in a code block, one subsection per component, tables for enums and mappings, a short code excerpt). 4 The central design insight, given its own section. 5 Experimental Results (setup, baseline table, loop table, qualitative differentiation). 6 A proposed architecture for the next version. 7 Scale. 8 Integrations. 9 Limitations and Future Work. 10 Conclusion. References (eleven entries, arXiv IDs given, mixed academic and industry sources).

Why it reads as research. The paper states a mechanism, names its inspirations with citations, reports numbers in tables with the setup spelled out, and has a limitations section that concedes real weaknesses (mock fidelity, judge dependence, single-turn only). The README and the paper are two doors into the same object: the README for people who want to run it, the paper for people who want to cite it.

What we copied

  • Root-level README plus a paper in markdown, both in the repo, cross-linked.
  • README opening: hero, one-sentence problem, transcript demo, "paste into Claude Code" quick start, cost table.
  • Paper skeleton: abstract, introduction with numbered core ideas, related work with per-subsection differentiation, system section with a pipeline diagram, a section devoted to the one central thesis, results as tables, limitations, references with arXiv IDs.
  • CLAUDE.md as the agent-facing map of the codebase.
  • MIT for code. CI on push and pull request.
  • GitHub topics, many and specific.

What dojo.md does not do (and a frontier-lab reader checks next)

  • No CITATION.cff, so no "Cite this repository" button and no machine-readable citation.
  • No tagged releases, no changelog, no DOI.
  • No arXiv version of the paper; the paper exists only in the repo.
  • No committed raw outputs: the results tables cannot be regenerated from files in the repo.
  • Model names in the paper are aliases (for example "Claude Sonnet 4.6"), not immutable snapshot IDs with access dates.
  • Prompts are in source code, not published verbatim as versioned artifacts.
  • No preregistration, no null model, no statistics beyond point scores.
  • No CONTRIBUTING, CODE_OF_CONDUCT, SECURITY, or issue templates.
  • The paper's results section is small (six scenarios, two models, one loop iteration) and the paper says so.

None of this is a criticism of the repo, which is a product with a paper attached. Our repo is a research artifact with a product attached, so the bar is different.

What we added

See research/01-publishing-practice.md for the evidence behind each item.

  • PREREGISTRATION.md, hashed and timestamped before any run, with a deviations log.
  • paper/ directory instead of a root paper.md: markdown source, refs.bib, a NeurIPS-style CHECKLIST.md, a figures folder, and a README on producing the arXiv PDF.
  • CITATION.cff, CHANGELOG.md, CONTRIBUTING.md, CODE_OF_CONDUCT.md, SECURITY.md, ROADMAP.md, AGENTS.md (with CLAUDE.md pointing at it).
  • Dual licensing: MIT for code, CC-BY-4.0 for paper and released data.
  • prompts/ with every prompt verbatim and versioned; run metadata records the prompt hash and the exact model snapshot ID.
  • experiments/ with one config per arm and committed raw outputs per run; make reproduce rebuilds every table without API keys.
  • design/ with one architecture decision record per load-bearing decision from the brief.
  • research/ with the full audit trail: the founding brief, the lineage note, and every agent sweep, unedited.
  • A public reproducibility corpus alongside the private primary corpus, so anyone can run the pipeline end to end.
  • A statistical null (label permutation) separate from the stimulus control (random pairing).