eCAADe 2026 // Miro Roman
STOP WRITING
LIKE IT'S 1995
Architectonic Writing
Nothing should be accepted as fixed, permanent, or ‘given’ — neither material conditions nor social forms.
If nature is unjust, change nature. If the format is unjust, change the format.
XF mutates, navigates and probes every horizon. // Laboria Cuboniks, 2015
The database is a designed artefact.
The AI constellation is a designed workflow.
But neither determines the argument —
the arguments determine them.
You push and rotate your constellations of AIs
around the main claims
until they hold:
sharp, clear, and directed.

XF refuses to mourn.
It constructs.
Paper #281

search0more

A Custom AI-Assisted Search Engine for 3D Architectural Assets
1 Probabilistic text and image encoding and search.
2 A specialised database serves as a critical node within a larger AI constellation.
Paper #290

Domain-Specific Libraries as Focus Lenses

for AI-Assisted Architectural Writing
1 LLMs write generic architecture.
2 Libraries sharpen the focus.
0x01 // ZERO — The Infrastructure
search0more
19,123 assets.
Text search. Image search.
Runs on a $20/month server.
PBR materials, 3D models, curated raster images. Latent Semantic Indexing for text. VGG19 + Self-Organizing Maps for images. Plugs into Blender. No GPU at query time.
11,602 unique stems
200 LSI dimensions
50 × 50 SOM grid
10,000 training iterations
VGG19 // ImageNet weights // frozen
3 × 2 patch grid per thumbnail
confidence threshold ≥ 0.333
Flask + AWS S3 + DynamoDB
0x02 // INTERRUPT — The Rupture
30 Jupyter notebooks.
A running application.
And a format you don't believe in.
If nature is unjust, change nature. If the format is unjust, change the format. The academic paper is not a neutral container — it is a disciplinary apparatus that systematically erases process, friction, and the actual conditions of production. To accept it uncritically is to naturalise a politics of erasure. Ours is a world that swarms with technological mediation, interlacing our daily lives with abstraction, virtuality, and complexity. XF constructs a feminism adapted to these realities — a feminism of unprecedented cunning, scale, and vision. No more futureless repetition on the treadmill of capital. No more reification of the given masked as critique. Our future requires DEPETRIFICATION.
0x03 // TRAP — The Constellation
Not one tool.
An ensemble.
Each member unreliable in its own specific way — which is exactly the point. The constellation is not a hierarchy. It is a network of partial capacities, strategically assembled. No single node holds authority. The human is not above the machine; the machine is not below the human. The architecture of the collaboration is the argument. XF seeks to strategically deploy existing technologies to re-engineer the world. The real emancipatory potential of technology remains unrealised — not because the tools are wrong, but because the constellations are.

Miro

Human Composer

Built the project. Holds taste, judgment, disciplinary intention. Decides everything.

Claude

Writing Partner

Long context. File access. Read the eCAADe proceedings. Learned the discourse. Drafts, proposes, cross-references.

Claude Code

Technical Analyst

Terminal agent. Reads the codebase directly. Extracts exact parameters. Traces every pipeline step.

ChatGPT

External Reviewer

Different model, different biases. Fresh eyes on every draft. Catches what the others miss.

Figure 1: The Network
Figure 1 // The Network: Humans, AIs, Libraries, and Orchestration Tools
Figure 2: The Constellation
Figure 2 // The Constellation: A plug-and-play assembly of actors
Paper #290
“The library exists within a constellation of actors: LLMs, image generators, search engines, data repositories, orchestration platforms, and human composers, each with specific capacities.”
0x04 // PARITY — The Two Inputs
Context is everything

search0more codebase

30+ notebooks // Flask app // trained models // figures
↓ Claude Code
= PRECISION

eCAADe 2025 Proceedings

Volume 1 + 2 // 200+ papers // 348 MB
↓ Claude Cowork
= POSITION
Paper #290
“Rather than relying on the LLM's own training data, which reflects the statistical average of the internet, the library introduces curated, intentionally partial, and disciplinarily dense material that conditions the LLM's field of articulation.”
0x05 // ADJUST — The Concept
Domain-Specific Libraries
as Focus Lenses
Without context:
“Generic plausibility — a kind of statistically averaged common sense that sounds correct but carries little disciplinary weight.”
With context:
Disciplinary specificity. Positioned writing. The right references, the right tone, the right intellectual framework. The library is not a filter — it is a lens. It does not subtract. It focuses. It is through, and not despite, our alienated condition that we can free ourselves from the muck of immediacy. The library is the technomaterialist instrument — it refracts generic plausibility into disciplinary precision.
Paper #290 // citing Bühlmann, Hovestadt & Moosavi (2015)
“If computation is literacy, then the way one structures queries, builds search environments, curates databases, and orchestrates prompt sequences is itself an architectural act.”
0x06 // CARRY — The Alien Territory
Working with
what you don't know
The parasite enters through the noise.
Not the signal — the interference.
348 MB of proceedings you haven't read. 200+ papers in a discourse you're entering sideways. The proper academic move would be two years of careful reading, a literature review, a positioning paper, then maybe a submission. The improper move: feed it all to an AI and start arguing.
Three weeks later you know the discourse — not because you sat in a library, but because you spent three weeks fighting with machines about where your work belongs. The knowledge came from the friction, not the reading.
Serres would call this the third position: neither inside the discipline nor outside it, but on the passage between. The navigator who learns the territory by crossing it, not by mapping it first. Alienation as method. The outside as productive position. Freedom is not a given — and it’s certainly not given by anything ‘natural.’ The construction of freedom involves not less but more alienation; alienation is the labour of freedom’s construction.
0x07 // OVERFLOW — The Process
Five Phases
1

Analysis

Claude Code reads the codebase. Extracts every parameter. Tells you things you forgot about your own code.

2

Discourse Learning

Claude reads 200+ papers from the eCAADe proceedings. Maps the community. Who cites whom. What vocabulary. What counts.

3

Drafting

Claude proposes. Miro accepts, rejects, redirects. Iterative. Argumentative. The good kind of friction.

4

Cross-Review

Sections go to ChatGPT. Different model catches different weaknesses. Then back. Then forward again.

5

Polish & Submit

Double-blind formatting. Anonymization. References. Figures. Final read. Submit to eCAADe 2026.

Paper #290
“The constellation is not fixed. Actors can be added, removed, rearranged, or substituted; the design of the constellation is itself a design act.”
0x08 // FEEDBACK — What It Actually Looks Like
The Dynamics
This isn't a clean pipeline. It's a conversation. Arguments. Wrong turns. Redirections. The mess is the method.
Claude Code
Reads 30 Jupyter notebooks, the Flask app, the search service, the trained models. Tells me things I forgot about my own code. “Your SOM grid is 50×50, trained for 10,000 iterations, with a confidence threshold of 0.333.”
Miro → Claude
Uploads 348 MB of eCAADe proceedings. “Read this. Learn the discourse. Tell me who cites whom, what vocabulary matters, what counts at this conference.”
Claude
Maps 200+ papers. Identifies clusters. Finds relevant references. Proposes a positioning strategy. Drafts introduction and related work.
Miro → ChatGPT
Copies Claude's draft. Pastes it into ChatGPT. “Comment on this. What's weak? What doesn't land?” Different model, different blind spots, different strengths.
Miro
“No. Not like that. Rewrite. Stronger. More precise. Drop the hedging.” Rejects. Redirects. Accepts one paragraph out of ten. Moves things around. Argues.
Claude
Unpacks the .docx XML. Fixes every font to Arial. Adds figure cross-references. Checks that all 42 references across both papers have valid URLs. The boring work that eats weeks.
The Loop
This cycle runs dozens of times. Not a prompt-and-done. A sustained editorial relationship across weeks. The paper emerges from the friction.
Result: 2 papers submitted. Both accepted. 45 references. 8 figures. Zero hallucinated claims — because every technical detail was extracted from real code by a real agent reading real files.
0x09 // DEPETRIFY — The Real Problem
The Paper Is
Junkspace
Junkspace was what remained after modernization ran its course. The scientific paper is what remains after knowledge production ran its course. A container that outlived its contents. A format designed for paper, binding, postage — none of which exist anymore. It survived the way parking lots survive after the shopping mall closes. We will tear down melancholy and illusion, the unambitious and the non-scaleable, and the libidinised puritanism of Nature as an un-remakeable given. The format is not sacred. It is REMAKEABLE.
The project:
30 notebooks full of dead ends and surprises. A running search engine. 19,123 assets. The moment when VGG19 features clustered concrete textures next to brutalist photographs and you thought: yes. Design decisions made at 2 AM. Code commits. Refactors. A Blender plugin that actually works.
The format:
8 pages. Double-blind. Template margins. “This paper presents...” “Results demonstrate...” Remove the personality. Flatten the process into methods-results-conclusion. Add a limitations section where you apologize for doing anything at all.
The good news:
If AI can compress your work into 8 pages, it can also decompress it back out.
THIS IS NOT FAKE
It's a render.
0x0B // MUTATE — The Argument
The Render
The project is real. The code is real. The application runs online.
The AI doesn't produce the research. It helps you render it — translate the same work into different formats, for different audiences, without losing what matters each time.
search0more
↓ ↓ ↓ ↓

Paper

eCAADe 2026 // 8 pages
→ scientists

Lecture

this page // also made with AI
→ educators

Studio

design course brief
→ students

Demo

live running app
→ practitioners
Same project. Same codebase. Same design decisions.
Four outputs. None of them fake. All of them partial. That's what rendering means.
Paper #290
“Authorship shifts from manually producing every sentence toward selecting datasets, framing queries, orchestrating relations, and critically editing outcomes. This mode of working is closer to architectural practice than generic chatbot use.”
Paper #290
“Writing becomes not a shortcut to generic text, but an architectural practice of selection, positioning, and composition.”
The author is not the one who types.
The author is the one who decides.
And if scientific writing kills what’s interesting about your work — render it again. Differently. For the people who actually care. Alienation is not the problem. Alienation is the engine. XF mutates, navigates and probes every horizon. It is a feminism of unprecedented cunning, scale, and vision. Nothing should be accepted as fixed.
Miro Roman // UIBK // House of Coded Objects