RFC 0002 — Publishing newblender discovery as a research series

Note

Generated from docs/rfc/0002-research-publication.md.

• Status: Open (proposal received, nothing decided)
• Opened: 2026-10-07
• Author: mbaneshi (with Claude Code)
• Depends on: RFC 0001, R5 (four-pass study method), the S01 brief
• Discussion: to be linked once the GitHub repo exists

1. Proposal (condensed from an outside review of the S01 brief)

The S01 brief reads as a research report / lab notebook / mid-run study, not yet an academic paper. Its data set is what separates it from an opinion piece about “Blender is bad for AI”:

• 78 production .blend files (~2 GB in the sample, 4.2 GB in ~/newblender-data/ overall)
• 2,498 registered operators; 1,605 context-dependent; 190 needing mouse/keyboard; 138 with no direct invocation; 320 modelling operators
• topology, modifier and Geometry Nodes statistics; a cited professional workflow; developer design threads; source tracing

Reframe. The research problem is bigger than Blender:

Before a professional creative application can become agent-native, we must learn how professionals actually use it, what production artefacts really contain, and which parts of its API still assume a human behind a mouse and keyboard.

Blender becomes the case study; the idea is agent-native creative software.

1.1 Series, not a single paper

• NEWBLENDER Research — S01: Modeling
  • S01 Mid-Run Report: the current brief, published as-is
  • S01 Final Report: after pass 3 (trace) is complete
  • Paper: extracted from the final report
• Later stages follow the same shape (S02 Animation, S03 Materials, …).

1.2 Candidate titles

• NEWBLENDER S01: An Empirical Study of Blender Modeling Workflows, Production Artifacts, and Human-Centric APIs
• Toward Agent-Native Creative Software: An Empirical Study of Blender’s Modeling Workflow and API (reviewer’s preference: Blender is the case study, the thesis is the headline)

1.3 Paper skeleton

1. Abstract. Problem (creative software is built around human interaction) · question · method (four passes) · results · implication.
2. Introduction. Not “I want to rewrite Blender”, but what would need to change before a professional creative application can become agent-native.
3. Research questions.
   • RQ1 What operations make up professional Blender modelling workflows?
   • RQ2 What structures actually occur in production .blend files?
   • RQ3 Which modelling operations depend on human UI context?
   • RQ4 Where does the programmatic interface encode assumptions about human interaction?
   • RQ5 Which existing abstractions already suit agentic execution?
   • RQ6 What architectural changes would make modelling more agent-native?
4. Methodology. Workflow → production files → operator/API surface → source → design rationale → agent-native principles.
5. Dataset. Formal counts and provenance; the stylised-film sampling bias moves into Threats to Validity.
6. Results. Operator-surface table (total / context-dependent / input-dependent / no direct invocation / modelling).
7. Production reality. 91% of sampled mesh objects carry modifiers; Subdivision Surface is 54% of modifiers. Claim: professional modelling is substantially more procedural and declarative than a naive replay of UI actions suggests.
8. Human-centricity. Knife, context, mouse/keyboard, Edit Mode, undo, operator redo, mesh copying, UI state; formalise the term human-centric API surface.
9. Agent-native design implications. Human-centric (context-dependent, imperative, stateful, interactive) vs agent-native (context-independent, declarative, inspectable, reversible, composable, verifiable); capability cards as the bridge.

1.4 Repository shape

newblender/
├── research/
│   ├── README.md
│   └── S01-modeling/
│       ├── protocol.md  dataset.md  methodology.md
│       ├── mid-run-report.md  final-report.md
│       ├── results/  evidence/  citations.md
├── data/
│   ├── manifests/   # source, licence, hash per file
│   ├── derived/
│   └── statistics/
├── experiments/
├── docs/
└── src/

Raw production files stay out of git: file → manifest → source/licence → hash → derived statistics. Reproducible without a 2 GB asset dump.

1.5 Publication channels

Channel	Role	Rating
GitHub repository	Code, data manifests, method, evidence (canonical)	★★★★★
GitHub Pages	Readable, permanent report site	★★★★★
Zenodo	Citable release with a DOI	★★★★★
arXiv	Public technical paper, citing GitHub + Zenodo	★★★★
Hugging Face Papers/Repo	Only if the AI-agent tooling angle grows	★★★
ACM / SIGGRAPH / CHI	Peer review, later stage	★★★★★ (later)

1.6 Intellectual architecture

NEWBLENDER → Agent-Native Software
  ├── S01 Modeling ── workflow · artefacts · API surface · source · rationale
  ├── S02 Animation
  └── S03 Materials
        ↓
  Agent-native design principles → capability model → new Blender architecture

2. Open questions

• Q1 — Publish at all, and when? Mid-run report now, or wait for the S01 final report? Does going public early cost anything (scooping, locking in claims before pass 3)?
• Q2 — Title and thesis. Blender-centred title or “Toward Agent-Native Creative Software”? The second makes a bigger claim that S01 alone may not carry.
• Q3 — Repo restructure. Adopt research/S01-modeling/ + data/manifests/ now, or keep docs/discovery/ and map it to the paper shape only at publish time? (RFC 0001’s link structure depends on the current paths.)
• Q4 — Dataset licensing and redistribution. Per-file licence audit of the 78 files (Blender Studio CC-BY variants vs others). Manifest + hash only, or redistribute derived statistics? Who checks?
• Q5 — Evidence standard. The brief is evidence level L1 (documented and traced, nothing benchmarked). What level must claims reach before a paper: runtime probes (brief §8), replication script, re-run on a second Blender version?
• Q6 — Visibility. Repo is local-only today (no remote). Public from day one, or private until the first release? This also decides where this RFC’s Discussion lives.
• Q7 — Channels and order. GitHub → Pages → Zenodo DOI → arXiv as proposed? arXiv needs an endorser for cs.GR/cs.HC; is that available?
• Q8 — Authorship and AI disclosure. How to credit agent-performed passes (R5: the agent carries all passes and first drafts) under arXiv/ACM authorship and AI-use policies.
• Q9 — Language. English canonical paper; Persian report/digests as a companion edition or not?
• Q10 — Next concrete step. The reviewer offers a paper-ready specification for S01 (title, abstract, RQs, methodology, dataset protocol, evidence standard, results schema, bibliography) without rewriting current findings. Take it as the first deliverable?

3. Resolutions

None yet.