MuNa/paper/paper.md

Two Languages, Two Tracks: A Live Public-Advocacy Deployment of LLM-Mediated Pattern Language Synthesis

Bruno Postle¹ and Nikos A. Salingaros²,³

¹ Union Street Research, Sheffield, UK ² Department of Mathematics, The University of Texas, San Antonio, TX, USA ³ Thrust of Urban Governance and Design, Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China

Draft — target venue: Buildings (sequel to Postle, B.; Salingaros, N.A. "LLM and Pattern Language Synthesis: A Hybrid Tool for Human-Centered Architectural Design." Buildings 2025, 15, 2400).

---

Abstract

A prior paper (Postle & Salingaros, Buildings 2025, 15, 2400) introduced a method for turning a curated subset of Christopher Alexander's A Pattern Language into an experiential design narrative using a large language model (LLM), demonstrated on a hypothetical university building. This paper reports a live deployment of an extended version of that method in an active, contested public process: an independent counter-proposal for the new National Museum of Ecuador (Quito), written in response to public objection to the winning entry of an international design competition. We extend the original two-input chain (pattern subset + brief) with a third input, a form language — a project-specific grammar of climatic response, tectonics, material, color, and ornament, following Salingaros' theoretical framework of pattern language and form language as complementary, jointly necessary instruments. We generalize the bespoke prompting used to produce this project's form-language grammar into a reusable, placeholder-driven template, and validate that it generalizes to an unrelated typology, region, and climate. From a shared narrative root and a shared pattern subset and form-language grammar, the two authors produced two divergent elaborations — two narratives and two illustrated image sets (13 watercolor-register images; 25 photo-realistic images) — that converge on compatible design moves while diverging sharply in visual register, a divergence we attribute partly to a deliberate difference in requested image medium and partly to a difference in how much of the source material each image-generation prompt carried. We report informal peer reception of this work at a public symposium, including a substantive critique from a practicing architect that the method, as run, produced "flavors, not solutions" because no single coordinated floor plan underlies the independently generated images. We discuss this and other limitations, and note — as one possible direction for future work, not yet attempted — that an explicit plan-layout generation stage could address it by mining spatial-adjacency information already implicit in the pattern subset.

Keywords: pattern language; form language; large language models; human-centered architecture; generative AI in design; participatory design; Christopher Alexander; Ecuador; museum architecture; public-advocacy design

---

1. Introduction

The dominant response of AI text-to-image tools to architectural prompts is to invent an attractive building from a short stylistic description — producing images that are slick, stylistically confused, and spatially and culturally arbitrary, because nothing upstream of the image constrains it. Postle and Salingaros [1] proposed an alternative: place image generation at the end of a structured chain rather than at the start, so that an LLM synthesizes an experiential design narrative from a curated, human-readable set of Alexander's [2] patterns before any image exists, and treats that narrative — not a mood-board prompt — as the instrument that generates and disciplines images. That paper demonstrated the method on a hypothetical university department building and used two independent LLMs to cross-check a claim about the pattern-generated design's likely effect on occupant productivity.

This paper reports two things the original paper could not: a live deployment of the method under real stakes, and a method extension that the deployment required.

The deployment: the Government of Ecuador ran an open international call (Stage 1 deadline 17 February 2026) for the architectural preliminary project of the new National Museum of Ecuador, on an approximately 13,000 m² site in Iñaquito, Quito. A winner was selected, but the winning design provoked sustained public objection over its failure to serve the people who will actually use it, and — at the time of writing — is on hold and may be abandoned. Independently of that competition, a small informal team, including both authors of this paper, used the method described in [1], extended as described below, to write and illustrate a complete counter-proposal: not a competition entry, but a public-advocacy document intended to influence the outcome of the debate while the winning scheme's fate is undecided. Unlike the earlier paper's case study, this deployment was not written for the purpose of demonstrating the method — the method was used because it was the fastest available way to produce a serious, human-centered alternative in the time available, and this paper is a report on what happened when it was used that way, in public, under scrutiny.

The extension: the original method combined a pattern subset with a project brief. For a museum expected to express Ecuadorian cultural identity as much as to organize human activity, a pattern subset alone was not sufficient — Alexander's patterns govern human/organizational relationships (how people arrive, gather, move, rest) but say little about how a building should look, or which regional material and ornamental vocabulary it should draw from. We therefore added a third input, a form language, following the theoretical distinction Salingaros sets out in "Two Languages for Architecture" [3]: a pattern language encodes rules for how human beings interact with built form, while a form language encodes geometric and tectonic rules for how matter is put together, and the two are jointly necessary — a form language that is not adaptive to human sensibility cannot connect to a pattern language, and a pattern language has no way to become a specific building without a form language. Section 3 reports how this form-language instrument was produced, and how we generalized its production into a reusable template.

The rest of the paper is a case study of what happens when this three-input method is run twice from a shared narrative root and shared pattern/form-language inputs, by two different authors, in public, and then subjected to informal peer critique. Section 2 gives the background of the competition and the specific, named failures the counter-proposal was written to avoid. Section 3 describes the method extension. Section 4 reports the worked example as two divergent tracks. Section 5 reports reception at a public symposium, including a substantive critique that the method, in its current form, does not yet produce a coordinated plan. Section 6 discusses this and other limitations honestly. Section 7 briefly sketches one possible direction for future work suggested by that critique. Section 8 concludes.

2. Background: The Museo Nacional del Ecuador Competition and Its Discontents

The National Museum of Ecuador holds the country's national collection — over 1.2 million objects and documents spanning roughly 12,000 years, across five reserves (archaeology, colonial and republican art, modern and contemporary art, a human-sciences library, and a historical archive) — of which the institution states that under 1% can currently be exhibited at once, for lack of space. The Stage 1 open call sought a preliminary architectural project for a new building of at least 25,000 m² on an approximately 13,000 m² plot at the corner of Av. Eloy Alfaro and Av. de la República in the Iñaquito parish of Quito, diagonally across from La Carolina Park — one of Quito's most-visited public spaces, itself drawing on the order of 400,000 visits a month — with direct access to the La Carolina Metro station.

A winning scheme was selected, but has since drawn considerable public objection. Independently of this paper's authors, a critique of the winning proposal was recorded and circulated, analyzing it against human-centered design and pattern-language principles, with eye-tracking-derived visual-attention data supporting several of its claims. That critique named specific, concrete failures rather than general aesthetic complaints: an entrance that visual-attention scans showed users could not locate (98% of attention going to non-entrance openings and parked cars); blank facades assessed as generating anxiety rather than welcome; and other legibility and human-scale failures. We mined this critique for named, evidenced failures and mapped each to the pattern(s) in our subset that directly counters it (e.g., illegible entrances countered by Main Entrance / Entrance Transition / Entrance Room), so that the counter-proposal's design decisions would explicitly and traceably steer clear of documented problems, not just gesture at being different in tone.

The counter-proposal — Casa de la Memoria Andina, in Postle's track; Casa de la Memoria Ecuatoriana, in Salingaros' — is the product described in the rest of this paper.

3. Method Extension: A Third Input — the Form Language

3.1 Two Complementary Languages

Salingaros' theoretical framework [3] holds that a pattern language and a form language are distinct and jointly necessary. The pattern language supplies the organizational code: which spatial relationships make a building legible and humane — where entrances go, how circulation reads, how outdoor rooms catalyze use. The form language supplies the cultural code: how the building meets the ground, how walls and roofs are built, which materials, colors, and ornament belong to the place, and where they are permitted to concentrate. A pattern-generated design without a form language has no way to decide what it is made of or what it looks like; a form language without a pattern language risks producing a building that looks locally authentic but is organizationally illegible or hostile to use. The original paper [1] used only the first instrument. This project required both.

3.2 The Ecuadorian Form Language Grammar

For this project, Salingaros produced a project-specific form-language grammar, the Ecuadorian Form Language Grammar, developed after discussion with a Quito-based designer and self-builder and synthesized by an LLM from open-source reference imagery. Rather than a set of historical facades to imitate, it is written as a generative vocabulary: a stratified tectonic sequence from dark Ecuadorian andesite at the plinth, through thick limewashed tapial and adobe at the lower and middle levels, to lighter bahareque and timber at the crown, under red clay tile roofs with deep aleros; a spatial lexicon of named local elements (atrio, zaguán, corredor, balcón corrido, mirador, canales, rejas); a restrained base color system ("Quito White & Andesite") with a wider selectable palette; and an ornament system concentrating geometric chakana and diamond-course motifs at base, threshold, and crown, plus a single permitted figurative retablo episode — with explicit prohibitions against superficial historicism, arbitrary facade color patches, and decoration applied without tectonic reason. Where the pattern language tells the building how to work, the form grammar tells it how to belong.

3.3 Templating the Form-Language Step

The Ecuadorian Form Language Grammar was originally produced by a bespoke, two-step prompting process specific to this project: an initial exploratory prompt asking an LLM to research vernacular and historical form across the region and propose a generative grammar rather than a style to copy, followed by a second prompt narrowing scope to the specific site and climate while retaining the wider regional vocabulary. This process was not, as originally run, reusable — it existed only as a conversational transcript specific to Quito.

We generalized it into form-language-prompt-template.md, a single-shot prompt template with placeholder fields for typology, region, site/climate, and historical strata, that collapses the two-step process into one prompt and produces a grammar document in the same structural shape (tectonic sequence, spatial lexicon, composition syntax, color and ornament systems) as the Ecuadorian worked example — analogous to how the web-based pattern-subset tool in [1] generalized manual pattern curation into a reusable instrument.

To check that this generalized, rather than merely described, we ran the template a second time on a project unrelated to Ecuador: Harbor House, a travellers' inn intended to relieve pressure on homeless services, sited in Sheffield, South Yorkshire — a different typology, region, and climate, produced by the paper's first author. This second run also used an LLM configuration with real web search available, so climate figures and precedent references were checked against live sources and cited inline rather than drawn purely from training-data recollection, demonstrating the template is compatible with a stricter, source-checked mode of use as well as the original's. The resulting document (templates/examples/sheffield.md) follows the same section shape as the Ecuadorian grammar while producing an entirely different vocabulary — dark gritstone and coursed rubble rather than andesite, slate and Yorkshire vernacular roof forms rather than red clay tile — which is the evidence that the template captures a reusable shape, not a repackaging of Ecuador-specific content.

4. Worked Example: Two Tracks From a Shared Root

4.1 Shared Inputs, Including a Shared Narrative Root

Both tracks reported below share the same project-specific pattern subset of 55 patterns from Alexander, Ishikawa, and Silverstein [2], curated using the APL-Companion web tool described in [1] and prepared for this project by Salingaros, and the same Ecuadorian Form Language Grammar (§3.2), condensed brief, and named-failure critique (§2). But the two tracks are not two independent runs from those inputs alone: they also share a common narrative ancestor. Salingaros' first prompt to an LLM (see Data and Materials Availability) combined the pattern subset, form language, and brief into an initial design narrative, Casa de la Memoria Andina. This document is the root both tracks below grow from, not merely a shared set of abstract rules: Track A's shared "chassis" document (its own set of fixed invariants — site data, the material/color/ornament systems, named circulation realms, and where each part of the collection lives) adopts it as primary source, and Track A's general-visitor narrative reproduces its prose almost wholesale (§4.2); Track B's final narrative is this same document merged with a second, independently drafted narrative from a different LLM (§4.3). The two tracks are consequently better described as divergent elaborations of a shared root than as fully independent runs — a distinction that matters for how the convergence reported in §4.4 should be read.

4.2 Track A: Four-Stakeholder Narrative Synthesis

Following the narrative-writing approach of [1] §4.2, Postle's track wrote four separate experiential narratives, one per stakeholder viewpoint (general visitor, museum staff, school group, and indigenous/community stakeholders whose heritage the collection represents). Of these, the general-visitor narrative is substantially the shared root narrative described in §4.1, reproduced with only the minimal adjustments needed for consistency with the settled chassis; the other three were written independently against the same chassis and pattern/form-language inputs, without directly reusing the root text. All four were cross-checked against the shared chassis document so that a detail asserted by one stakeholder's narrative (a material, a color, a named place) is never contradicted by another's, then combined in full, not condensed, into a single public presentation document, deliberately without pattern numbers or grammar codes, since it is written to be read by the public, not by practitioners. Thirteen illustration prompts were written, each self-contained (a shared style guide, distilled from the form-language grammar into plain descriptive language, pasted literally into every prompt, not cross-referenced — rather than the source pattern-subset or form-language documents themselves), and generated externally via Midjourney, producing images in a warm architectural watercolor and gouache register, deliberately requested — a painted presentation drawing, not a photo-realistic render — with lighting, weather, and a consistent cast of figures held constant across the set.

4.3 Track B: Single Merged Narrative

Salingaros' track produced a single narrative, Casa de la Memoria Ecuatoriana, by merging the shared root narrative (§4.1) with a second, independently drafted narrative from a different LLM, explicitly instructed to combine the strongest elements of each against the same pattern subset and form language. It was then illustrated with 25 images from ChatGPT, requested and produced in a markedly more photo-realistic register, with no watercolor or illustrative-medium constraint applied. Salingaros reports, in his own first-person account of the process (see Data and Materials Availability), an iterative, natural-language critique loop run against these images: successive rounds were rejected and corrected in plain language rather than by editing pixels — a patio that read as faux-antique with trip hazards was regenerated with flush, level pavers; an entrance that did not read as a front door on approach was replaced by an axial view with a projecting five-bay portal; a building raised too far above the street was brought down to sidewalk grade; monochromatic tiled columns foreign to the context were regenerated with patterned azulejos; and overly white interiors, arbitrary facade color patches, and modernist metal railings were each identified and revised, each correction traced back to an explicit violation of the pattern subset or form grammar. This account is self-reported rather than independently logged; the delivered image set offers circumstantial, not conclusive, support — its file numbering skips several indices and includes revised variants for at least two images, consistent with an iterative reject-and-regenerate workflow without proving one. Each reported cycle of criticism and regeneration took minutes, not days.

Separately, and independent of any explicit request, each image prompt Salingaros wrote for ChatGPT included the full pattern-subset and form-language documents inline, rather than a distilled summary, while Track A's Midjourney prompts (§4.2) were shorter, LLM-distilled, scene-specific prompts referencing only the condensed style guide. This difference in prompt strategy — not only the difference in requested medium — is discussed as a candidate explanation for the image-register divergence in §4.4 and §6.3.

4.4 Convergence and Divergence

Table 1 summarizes the two tracks.

	Track A (Postle)	Track B (Salingaros)
Narrative structure	4 stakeholder-viewpoint narratives, combined in full	1 merged narrative
Relation to shared root narrative	visitor narrative = root text; other 3 written independently	root text merged with a second, independently drafted LLM narrative
Output document	docs/counter-proposal.md	docs/3.Casa-de-la-Memoria-Ecuatoriana-*.txt
Image generator	Midjourney	ChatGPT
Image count	13	25
Image register (requested)	watercolor / gouache illustration	no medium constraint; rendered photo-realistic
Image-prompt content	short, LLM-distilled style guide per scene	full pattern-subset + form-language documents inline per image
Consistency mechanism	shared chassis document (invariants)	iterative natural-language critique loop (self-reported, §4.3)

Because Track A's visitor narrative and half of Track B's merge input share the same root text (§4.1), part of the convergence between the two tracks is expected rather than independently earned, and should not be over-read as proof that the shared instruments alone drive agreement. The more meaningful evidence for that claim comes from the material that was not shared: Track A's other three stakeholder narratives, Track B's second, independently drafted merge input, and — most significantly — the two entirely separate image-generation stages, run by different authors with different generators and no shared images. Across that genuinely independent material, both tracks still arrive at the same high-level design moves: a stratified material sequence from stone base to timber crown; an atrio / zaguán / patio arrival sequence; a legible, projecting entrance portal; warm-toned interiors with restrained concentrated ornament; and circulation organized around named realms rather than corridors. This is evidence, from a smaller and less controlled sample than a fully independent two-track experiment would provide, that the pattern subset and form-language grammar meaningfully constrain the output beyond what shared text alone would explain.

The two tracks diverge sharply, however, in visual register: Track B's images are, in the authors' own judgment, markedly more photo-realistic and — subjectively — more visually compelling than Track A's watercolor/gouache-register images. This judgment has not been evaluated externally on a like-for-like basis: the symposium reported in §5 discussed Track B's images only, and Track A's images were not presented there, so it should be read as the authors' own assessment, not as corroborated outside reception. We attribute the divergence to at least two factors, neither yet isolated from the other or from any residual generator-quality difference: first, it is partly a deliberate, requested difference — Track A's prompts explicitly specify a watercolor and gouache medium, while Track B's prompts carried no illustrative-medium constraint and rendered photo-realistically by default; second, as reported in §4.3, Track B's prompts carried the full pattern-subset and form-language documents inline, while Track A's carried a short, distilled style guide, and the authors' working hypothesis is that ChatGPT's greater tolerance for long, dense prompts let it resolve more of the form language's specific vocabulary directly into each image, where Midjourney's shorter prompts left more to the generator's own defaults. This is a hypothesis, not a verified finding (see §6.3).

5. Reception: Informal Peer Engagement at the July 2026 Symposium

On 23 July 2026, this counter-proposal project was discussed at a public symposium including Salingaros, Nir Buras (Classic Planning Academy), Pablo Álvarez Funes, and others, walking through Track B's narrative and image process. We treat this as informal peer engagement, valuable precisely because it included substantive disagreement rather than only endorsement.

Participants praised the images' consistency of "spirit" — not pixel repetition of a single scheme, but a recognizable design language holding across independently generated views. But a practicing architect on the panel, Pablo Álvarez Funes, raised a critique that is the paper's most important piece of reception data: "this Nikos did not create a plan... we need to give a response to the program and ensure that all the spaces work coordinately." He characterized the current output as a set of well-crafted individual rooms rather than a linked building, adding that close inspection of the images — "especially at the corners" — reveals AI hallucinations, and summarized bluntly: "these are images. These are flavors. These are not solutions." The same discussion noted the continuing, non-optional role of licensed architects in taking such material to floor plans, elevations, sections, and mechanical/electrical/plumbing/structural coordination before anything could be built.

This critique is correct as stated, and it is not new information to the authors — it names precisely the gap described honestly in the original paper's limitations section [1] and inherited unchanged by this extension: the method produces experiential narratives and illustrative images of a design, not a coordinated plan. What the symposium adds is an unsolicited, on-the-record confirmation from outside the authors' own team that this gap matters. We return to it in §6.1 and note one possible response, briefly, in §7.

6. Discussion: Drawbacks and Open Questions

6.1 No Coordinated Floor Plan

Because each narrative and each image is generated independently against shared rules rather than against a shared drawing, spatial claims made across different illustrations of the same design are not guaranteed to reconcile. A concrete instance from this project: the narrative describes the entrance portal as composed in "a rhythm of five bays," but the corresponding illustration shows three arches. Rather than regenerating the image to match the text or rewriting the text to match the image, we treated the narrative as the design intent of record — it is the instrument the pattern subset and form grammar were synthesized into, and is where every design claim is traceable — and loosened only the image's caption to avoid asserting a bay count the image does not show. This is a defensible correction for a single instance, but it does not scale: it works because a human noticed the discrepancy and had a clear rule (text over image) for resolving it, not because the method prevents such discrepancies from arising.

6.2 No Mechanism for Genuine Public Participation

The method as deployed here produces material intended to influence public debate, but has no built-in mechanism for the public to actually participate in shaping it beyond viewing and reacting to finished narratives and images — e.g., informal social-media voting on generated images. Symposium participants warned that such unfiltered voting degenerates without curation (toward whichever image is most immediately eye-catching rather than most sound), which is a caution worth taking seriously if this method is to be used for participatory design rather than only advocacy communication.

6.3 Partly-Explained Divergence in Image Register

As reported in §4.4, Track B's photo-realistic images are, in the authors' own assessment, more compelling than Track A's watercolor-register images — a judgment the authors have made themselves, not one corroborated by outside reviewers evaluating both sets side by side. We attribute the divergence partly to a deliberate difference in requested medium (photo-realistic vs. watercolor/gouache) and partly to a difference in image-prompt content (the full pattern-subset and form-language documents fed inline to each Track B prompt, versus a short distilled style guide for each Track A prompt), on the working hypothesis that ChatGPT's greater tolerance for long, dense prompts let it resolve more of the form language directly into each image. Both factors are plausible and neither has been isolated from the other or from any residual difference in generator capability — a controlled comparison (same narrative text, same requested medium, both generators) would be needed to separate them, which is outside the scope of this paper.

6.4 Unsettled Questions of Authorship and Structural Readiness

As with the original method [1], these outputs are concept studies, not architecture: structure, egress, conservation environments, accessibility compliance, and budget still require conventional technical development, and questions of authorship and intellectual property in AI-assisted imagery remain unsettled in general, not resolved by anything specific to this project.

7. Possible Future Work

The critique in §5 points toward one possible direction for extending the method, which we note here briefly without committing to it or having attempted it: an explicit plan-layout generation stage, inserted between the pattern-subset/form-language inputs and narrative generation, that would make the spatial and adjacency relationships already implicit in the pattern subset (building complexes, wings, connected buildings, circulation realms) explicit as a shared layout for narratives and images to be checked against, rather than each independently inventing one. This is a proposal in prose only — no prototype has been built for this paper, and we do not treat it as the method's necessary next step so much as one plausible answer to a real critique among others an architect might reasonably prefer.

A separate, lower-priority open question is the image-register divergence of §6.3: a controlled comparison isolating requested medium from prompt-content length would help settle how much each contributes, but is not required to support this paper's central claims.

8. Conclusion

This paper reported a live deployment of an extended version of the LLM pattern-language synthesis method introduced in [1], used in earnest, in public, under real contested stakes, rather than as a hypothetical case study. The deployment required a genuine method extension — a third input, the form language, following Salingaros' theoretical distinction between the human/organizational and geometric/tectonic languages of architecture — which we generalized from a bespoke, project-specific prompting process into a reusable template and validated on an unrelated project. Run by two authors from a shared narrative root and shared pattern subset and form language, the method's two divergent elaborations converged on compatible design moves in their genuinely independent material, while producing a qualitatively different image register, a difference we can partly, though not yet fully, account for. Presented to informal peer scrutiny, the work drew a substantive critique, not only praise: that the method, as deployed, produces images without a coordinated plan behind them. We agree with that critique and take it seriously; §7 sketches one possible response without committing to it as the method's necessary next step. Design knowledge, once written down as readable rules rather than embodied only in a single expert's intuition, can be executed, checked, criticized, and revised at the speed of conversation — but a pattern language and a form language, however well executed, are not yet a plan, and this deployment's clearest lesson is that the method should say so honestly rather than let illustrative images stand in for one.

---

References

1. Postle, B.; Salingaros, N.A. LLM and Pattern Language Synthesis: A Hybrid Tool for Human-Centered Architectural Design. Buildings 2025, 15, 2400. https://doi.org/10.3390/buildings15142400
2. Alexander, C.; Ishikawa, S.; Silverstein, M. A Pattern Language: Towns, Buildings, Construction; Oxford University Press: New York, NY, USA, 1977.
3. Salingaros, N.A. Two Languages for Architecture. In A Theory of Architecture; Sustasis Press, 2006; revised 2014.

Data and Materials Availability

All primary materials referenced in this paper are project documents rather than independently published data, and are cited here by their path in the originating repository for traceability rather than as formal references:

• Pattern subset: ref/A Pattern Language Ecuador National Museum.md
• Ecuadorian Form Language Grammar: ref/Ecuadorian_Form_Language_Grammar.md
• Condensed brief and named-failure critique: ref/brief-condensed.md, ref/critique-failures.md
• Shared invariants ("chassis"): ref/project-chassis.md
• Shared root narrative and the prompts that produced it: docs/Casa-de-la-Memoria-Andina-National-Museum-Narrative-Kimi.txt, docs/salingaros-narrative-prompt.md
• Form-language prompt template: templates/form-language-prompt-template.md
• Template validation on a second project: templates/examples/sheffield.md
• Track A narratives and public document: narratives/01-visitor.md through narratives/04-community.md; docs/counter-proposal.md; illustration prompts, docs/illustration-prompts.md; images, docs/images/
• Track B narrative and images: docs/3.Casa-de-la-Memoria-Ecuatoriana-Design-Narrative-Museo-Nacional-del-Ecuador.txt; docs/images-salingaros/; process report, docs/From-Pattern-Language-to-Architectural-Image-Salingaros-revised.md
• Symposium transcript: ref/youtube-transcript-2026-07-23.txt
• Winning-scheme critique transcript: ref/youtube-transcript.txt