MuNa/paper/outline.md

Outline — Follow-up paper: plan-layout extension of LLM pattern synthesis

Working title: From Pattern to Plan: Extending LLM-Mediated Pattern Language Synthesis with a Form-Language Layer and a Live Public-Advocacy Deployment

Authors: Bruno Postle, Nikos A. Salingaros

Target venue: Buildings (sequel to Postle & Salingaros, "LLM and Pattern Language Synthesis: A Hybrid Tool for Human-Centered Architectural Design," Buildings 2025, 15, 2400 — ref/buildings-15-02400-v2.txt)

Status: outline only, agreed structure, no full prose yet.

1. Introduction

Recap the original pattern → narrative → image chain from the 2025 paper. Frame this paper as reporting a live, high-stakes public-advocacy deployment — not a hypothetical case study — plus a method extension: the Ecuador National Museum counter-proposal, an independent alternative developed while the winning competition scheme is on hold pending public objection.

2. Method extension: three inputs instead of two

The original method combined a pattern-language subset with a project brief. This project adds a third input, run in parallel with it: a form language, per Salingaros' "Two Languages for Architecture" (ref/Salingaros-Chapter-11-Two-Languages-for-Architecture.md) — the pattern language governs human/organizational rules, the form language governs geometric/tectonic rules (climate response, material, color, ornament, historical strata).

Inputs used:

• Pattern language: ref/A Pattern Language Ecuador National Museum.md (55 patterns, curated via APL-Companion)
• Form language: ref/Ecuadorian_Form_Language_Grammar.md
• Brief/critique: ref/brief-condensed.md, ref/critique-failures.md

3. Templating the form-language step

Salingaros' original form-language grammar was produced by a bespoke, non-reusable two-step prompt (docs/salingaros-form-language-prompt.md). This paper generalizes it into templates/form-language-prompt-template.md — a single-shot, placeholder-driven instrument (typology / region / site-climate / historical strata) that produces a grammar document in the same section shape, analogous to how the APL-Companion tool generalized pattern-subset curation in the original paper. Reusable on future projects outside Ecuador.

4. Worked example: two independent runs from shared inputs

Same pattern subset + same form-language grammar, fed into two separate LLM pipelines, run independently by the two authors, with divergent outputs:

• Postle track: four stakeholder narratives (visitor, staff, school group, community) → docs/counter-proposal.md, illustrated with 13 Midjourney images (docs/images/).
• Salingaros track: a single merged narrative, docs/3.Casa-de-la-Memoria-Ecuatoriana-Design-Narrative-Museo-Nacional-del-Ecuador.txt, illustrated with 25 ChatGPT-generated images (docs/images-salingaros/).

Report as a qualitative finding, not a controlled comparison: the two image sets differ markedly in register. Salingaros' ChatGPT images read as photo-realistic and are, by informal consensus among the authors and symposium participants (see §5), the more visually compelling of the two; Postle's Midjourney images are looser and more watercolour-like. The cause is not established — candidate variables include the generator itself, prompt phrasing, and how literally each generator resolves the same form-language vocabulary — and the paper should flag this openly rather than assert an explanation it hasn't tested. Given this, the paper's illustrative material should draw primarily on the Salingaros/ChatGPT image set, using the Postle/Midjourney set only for the explicit two-track comparison in this section, not as the paper's default visual register elsewhere.

5. Reception

The July 23 symposium (ref/youtube-transcript-2026-07-23.txt) as informal peer engagement, on the record:

• Praise for aesthetic consistency of "spirit" across the (Salingaros) images, without demanding pixel-identical repetition (~l.1295–1304).
• Substantive critique from a practicing architect (Pablo Álvarez Funes): "this Nikos did not create a plan" (l.1475); "these are images, these are flavors, these are not solutions" (l.1499); hallucinations visible on close inspection of corners (l.1496); need for licensed architects to turn the images into buildable floor plans, elevations, sections, and MEP/structural coordination (l.1214, l.1484).

6. Drawbacks

• Images are inspirational/illustrative, not derived from one coordinated floor plan — spatial claims across images aren't guaranteed to reconcile. Concrete instance from this repo: the entrance-portal image shows three arches against a five-bay text description; resolved by loosening the image caption rather than rewriting the design narrative (the narrative is the design intent of record).
• No mechanism yet for genuine public participation beyond unfiltered social-media image voting, which the transcript participants warn degenerates without curation.
• Open, unexplained variable: why one generator's output reads as more architecturally credible than the other's for the same inputs (§4) — noted as a limitation of the method's current image stage, not resolved here.

7. Next steps

Propose (not prototype, per author decision) plan-layout generation as a missing stage: mining the pattern subset's implicit spatial/adjacency relationships (e.g. project-chassis.md's named circulation realms) into an explicit layout layer that narratives and images could be derived from, rather than each being independently invented — directly answering the "this did not create a plan" critique in §5. See ecuador-0u5 for an exploratory prototype that may or may not land in time to fold into this section.

Secondary open question, lower priority: whether the image-register gap observed in §4 is attributable to the generator, the prompt, or something else — worth a controlled comparison (same narrative, same style-guide text, both generators) if time allows, but not required for this paper's core claims.

8. Conclusion