AI-Spiralism-Research-09-text-to-world-semantic-transformation.md

Full report

On this page

Executive Decision Summary

The ambition of "The Indoctrination Machine"—to transform a 3D geometric world simply by altering a single governing sentence—requires a rigorous architectural separation between linguistic interpretation and graphical execution. Relying on end-to-end generative AI models (such as latent-space text-to-3D diffusion) yields unpredictable visual hallucinations that frequently masquerade as systemic understanding, ultimately severing the semantic link between the visitor's authorship and the resulting artwork. The analysis indicates that achieving genuine, interpretable mesmerization necessitates deploying Large Language Models (LLMs) strictly as semantic routers. The LLM must parse natural language into a highly bounded, pre-authored structured recipe (a JSON schema), which subsequently drives a deterministic WebGL/Three.js rendering engine.

This procedural paradigm guarantees exact replayability, safe parameter validation, and explicit preservation of digital authorship. By revealing the mapping dimensions—rooted theoretically in the Osgood Semantic Differential and Attention Restoration Theory—directly to the user, the application avoids the pitfalls of an opaque "black box." Instead, it operates as a transparent theater of meaning-making. The proposed minimum viable implementation preserves a fully manual control mode as its baseline, utilizing the AI solely to propose slider configurations. This ensures the visitor retains ultimate editorial control, delivering the project's flagship promise without compromising browser performance, ethical transparency, or user consent.

Evidence Method and Reporting Discipline

Search Date: September 16, 2026

Substantive Search Strings:

1. "semantic differential" Osgood visual space dimensions factor analysis

2. "Metaphors We Live By" visual interaction design interface metaphors

3. "Attention Restoration Theory" Kaplan fascination visual geometry

4. "ShapeGlot" language grounded shape selection study

5. "PROV-O" provenance ontology W3C recommendation

6. Text2Mesh "3D mesh" text driven visual evaluation metrics

7. "3D-GPT" procedural 3D modeling language model schema

8. "hypnotic susceptibility" absorption visual art appreciation psychology

9. "visual metaphor" cross cultural variation empirical study HCI

10. "ChangeIt3D" shape deformation 3D latent space paper

Databases and Sites Examined:

arXiv (Computer Science and Linguistics repositories), W3C Standards Registry, IEEE Xplore, ACM Digital Library, PubMed Central, Three.js Official Documentation, and maintained GitHub repositories (ChangeIt3D, PartGlot, 3D-GPT).

Inclusion/Exclusion Decisions:

The research prioritized peer-reviewed empirical studies on semantic space, official W3C recommendations for data lineage, and recent (2022–2026) computational linguistics papers regarding 3D shape grounding. Literature detailing clinical hypnosis as a medical treatment was strictly excluded, shifting the focus toward psychological absorption and aesthetic appreciation. Non-reproducible opinion blogs regarding WebGL performance were excluded in favor of official technical documentation and large-scale rendering benchmarks.

Evidence Table

 

1. Survey of Semantic Visualization and Text-Controlled Design

The ambition to alter a complex 3D world by revising a single proposition relies heavily on resolving the tension between the inherent ambiguity of natural language and the strict mathematical requirements of geometric rendering. Current trajectories in generative artificial intelligence emphasize end-to-end models (such as latent-space text-to-3D diffusion) which ingest natural language and directly emit complete meshes, point clouds, or radiance fields10. While these models demonstrate high fidelity for discrete objects, they lack interpretable, granular control. When an end-to-end model is asked to alter a scene based on a subtle prompt modification, the entire topological space frequently collapses or hallucinates unrelated elements. This results in unpredictable spectacle rather than a semantic correspondence that a human observer can judge as meaningful21.

Conversely, procedural generation systems guided by LLMs separate linguistic comprehension from geometric execution. Architectures such as 3D-GPT or Code-as-Room utilize the LLM strictly as a reasoning engine, mapping semantic intent into a structured, machine-executable schema10. This schema is subsequently passed to a deterministic rendering engine. This paradigm ensures that the visual transformation is directly tied to authored, bounded variables, preventing random visual artifacts from masquerading as systemic comprehension. Furthermore, research in dynamic typography demonstrates that similar deterministic constraints are required when animating text; without vector-based structural preservation techniques (such as Bézier curve constraints), semantic deformations quickly render typography illegible13.

Empirical studies in shape differentiation, notably ShapeGlot and ShapeTalk, demonstrate that natural language can successfully map to fine-grained 3D geometries through "reference games," wherein a neural listener learns to distinguish a target shape from a distractor based on a linguistic description7. However, these frameworks largely focus on physical object parts (e.g., identifying that a chair has thinner legs). To map abstract, philosophical propositions like the doctrines in "The Indoctrination Machine," the system must lean heavily on conceptual metaphor, translating abstract cognitive states into spatial relationships utilizing established psychometric frameworks26.

2. The Mechanism of Fascination: Psychological and Aesthetic Constraints

The project mandates an experience that is mesmerising and provocative without resorting to covert conditioning, deliberate strobes, or clinical hypnosis. The design must therefore leverage the mechanics of psychological absorption and "soft fascination."

Attention Restoration Theory (ART) posits that specific environmental geometries possess the capacity to gently capture involuntary attention without demanding exhausting, directed cognitive focus5. Natural environments frequently achieve this through statistical fractals—patterns that repeat at multiple scales with slight variations6. Because visual processing is hierarchically distributed, environments containing balanced geometric complexity across multiple scales allow the human cognitive system to remain deeply engaged while simultaneously resting6. ART identifies four primary components necessary for this restorative fascination: Extent (a sense of vastness and coherence), Being Away (conceptual separation from mundane tasks), Fascination (effortless attention), and Compatibility (alignment with the user's intrinsic motivations)5.

In the context of the proposed artwork, mesmerizing visual geometry should mimic these statistical fractals to achieve Extent and Soft Fascination. The proposed "Monolith" and "Chorus" structures must explicitly avoid high-frequency flickering or over-stimulating contrast that triggers fatigue or violates accessibility standards (e.g., WCAG 2.2 Success Criterion 2.3.1 regarding photosensitive flashes)31. Instead, visual fascination can be achieved through slow, smooth, instanced repetition, where continuous micro-variations in a Three.js InstancedMesh transformation matrix create a sense of vastness and organic coherence without overwhelming the browser's draw-call budget19.

It is crucial to distinguish this aesthetic absorption from hypnotic susceptibility. While the personality trait of "absorption" strongly correlates with the ability to deeply appreciate visual art and willingly suspend disbelief, it remains a self-directed, conscious state of engagement32. The artwork's framing of "indoctrination" is explicitly theatrical. By allowing the user to view the mechanisms of change, explicitly edit the governing sentence, and control the motion state via strict opt-in gates, the experience maintains ethical transparency. It operates as an interactive exhibit demonstrating the power of semantics, rather than functioning as a psychological intervention.

3. Defining Candidate Visual Dimensions: The Visual-Semantics Dictionary

To map abstract, unstructured sentences to deterministic 3D scenes, the system must utilize a defined vocabulary of visual dimensions. These dimensions are deeply grounded in Charles Osgood's Semantic Differential, a psychometric technique that demonstrates human meaning is universally structured along three bipolar axes: Evaluation (Good-Bad), Potency (Strong-Weak), and Activity (Active-Passive)1.

The following Visual-Semantics Dictionary outlines 15 candidate mappings. Dimensions 1 through 5 are strongly supported by cross-cultural empirical findings regarding spatial metaphors. Dimensions 6 through 10 are derived from theories of visual psychology and Attention Restoration Theory. Dimensions 11 through 15 are explicitly authored artistic mappings designed specifically for the theatrical premise of "The Indoctrination Machine."

If empirical testing reveals that participants fail to reliably associate a shift from "authoritarian" to "democratic" with a visual shift from "Monolith" to "Chorus," this specific mapping dimension must be re-weighted, explicitly defined in the UI, or discarded entirely to preserve overall system legibility.

4. Sentence-Pair Case Designs and Semantic Parsing

To ensure the system interprets deep meaning rather than merely matching superficial tokens, it must appropriately handle tonal shifts, negations, and ambiguous wording. Simple keyword matching fails entirely on negation; a rule-based system might identify the word "complicated" in the sentence "The system is not complicated" and erroneously increase the scene's visual complexity. LLMs, when prompted to evaluate text along the Semantic Differential axes, can successfully synthesize these nuances36.

The following 12 original sentence-pair cases outline the transformation from a source sentence (State A) to a revised sentence (State B), detailing the targeted semantic shift and the corresponding visual execution.

1. Proposition Shift (Expansion)

• State A: "Every question has one answer." (Monolith archetype: A single massive structure, rigid grid layout, high symmetry, high centralization).
• State B: "Every question opens another possibility." (Unraveling archetype: The singular structure fractures into a branching, rhizomatic network, expanding outward).

2. Proposition Shift (Cohesion)

• State A: "We agree." (Chorus archetype: Multiple nodes pulsing synchronously in phase, unified color palette, harmonious curvature).
• State B: "We can disagree." (Chorus archetype: Nodes drift out of phase, colors diversify into analogous palettes, structural connections remain intact but visibly flex).

3. Temporal/Aspect Shift

• State A: "Unchanged." (Static mapping: Zero motion speed, high structural integrity, sharp angles, monolithic footprint).
• State B: "Still becoming." (Oscillation mapping: Instanced meshes begin a slow, breathing sine-wave translation; edges soften via erosion thresholds).

4. Volume/Space Shift

• State A: "I am contained." (Enclosure mapping: Dense clustering of instances, high opacity, inward-facing geometric normals).
• State B: "I am expanding." (Enclosure drops: Particles accelerate outward, the scale of individual nodes decreases while the overall scene bounds drastically increase).

5. Tone/Vulnerability Shift

• State A: "The structure is permanent." (Erosion mapping: Value set to 0\. High metallic material properties, solid bounds, maximal opacity).
• State B: "The structure is fragile." (Erosion mapping: Value set to 1\. Fragment shader noise increases significantly, introducing transparent voids; underlying wireframes become visible).

6. Power Dynamic Shift

• State A: "Authority flows from the center." (Centralization mapping: A single massive core, line segments physically tethering smaller satellite nodes to the center).
• State B: "Power is distributed everywhere." (Decentralization mapping: The center mass dissolves; line segments reconnect to nearest-neighbors, forming a peer-to-peer web).

7. Punctuation/Intent Shift (Command vs. Invitation)

• State A: "Do not look away." (Activity mapping: High. Aggressive focal depth pulling the eye to a single point, high saturation, sharp geometry targeting the camera).
• State B: "Look closer." (Activity mapping: Low. Soft fascination is engaged, smooth branching, depth of field softens to pull focus to intricate, slow-moving micro-structures).

8. Ambiguous Word Shift

• State A: "Everything is connected." (Order mapping: Symmetrical, geometric lattice, high regularity, even spacing).
• State B: "Everything is entangled." (Chaos mapping: Spline curves overlap chaotically, high geometric complexity, tangled branching without a clear spatial hierarchy).

9. Direct Negation (Challenging token matching)

• State A: "The path is clear." (Linearity mapping: Nodes align in a single descending/ascending corridor, high luminance, low erosion).
• State B: "The path is not clear." (Obstruction mapping: The linear corridor fractures, fog uniforms increase significantly to obscure distance, focal depth narrows).

10. Collectivism Shift

• State A: "We stand alone." (Isolation mapping: Distinct, unconnected pillars, spaced far apart beyond standard enclosure radii).
• State B: "We stand together." (Unity mapping: Pillars converge spatially, line systems wrap around the combined mass to demonstrate connectedness).

11. Fatalism Shift

• State A: "There is no escape." (Enclosure mapping: Concentric spheres trapping the camera viewport, low luminance, high potency).
• State B: "There is a way out." (Openness mapping: The concentric spheres break apart, a high-luminance light source appears in the distance, and geometry physically curves toward it).

12. Sensory Metaphor Shift

• State A: "Silence is golden." (Evaluation mapping: High. Warm lighting uniforms, slow oscillation, smooth curvature).
• State B: "Silence is heavy." (Potency mapping: High. Cool/dark lighting uniforms, geometry shifts to blocky, monolithic forms sinking downward on the Y-axis).

5. Architectural Paradigm: Generative AI as Procedural Composer

To strictly prevent the LLM from outputting unexecutable browser code, introducing Cross-Site Scripting (XSS) vulnerabilities, or failing silently, the system architecture must treat the LLM strictly as a data transformation layer. The visitor's text is analyzed by the LLM, which is prompted to output a strictly typed JSON object conforming to a predefined schema10. This JSON is then ingested by the client-side Three.js engine to manipulate the scene.

Proposed Structured-Recipe Schema

The JSON schema binds the 15 dimensions from the visual dictionary into normalized parameters, typically represented as floating-point numbers between 0.0 and 1.0.

 

 

 

JSON

{   "interpretation": {     "evaluation\_score": 0.8,     "potency\_score": 0.2,     "activity\_score": 0.5,     "semantic\_explanation": "The text implies growth and distributed potential rather than centralized control."   },   "geometry\_parameters": {     "archetype": "Chorus",      "complexity\_count": 5000,     "centralization": 0.1,     "branching\_factor": 0.8,     "enclosure\_radius": 0.9   },   "material\_parameters": {     "luminance": 0.7,     "saturation": 0.6,     "erosion\_noise": 0.2,     "curvature\_smoothing": 0.9   },   "motion\_parameters": {     "base\_speed": 0.1,     "oscillation\_amplitude": 0.3,     "direction\_bias\_y": 0.5   } }

Safe Parameter Validation

The frontend architecture must rigorously validate this JSON payload before execution. The archetype string must match an internal enumeration (e.g., \["Monolith", "Chorus", "Unraveling"\]). Numeric values must be clamped to safe ranges; for instance, the complexity\_count must be restricted to a maximum threshold (e.g., 10,000) to maintain the draw-call budget and ensure a stable 60 Frames Per Second (FPS) target across mid-range devices39.

Values passed directly to ShaderMaterial uniforms must be sanitized to prevent buffer overflows or shader compilation errors40. Because the LLM solely selects bounded scene actions and is prohibited from writing raw JavaScript or GLSL, arbitrary code execution is mathematically impossible. Furthermore, visual consent (motion permission) is handled strictly by the browser session state and is never encoded into, or overridden by, the LLM's recipe.

6. User-in-the-Loop Interpretation and Ambiguity

If the semantic transformation occurs entirely silently in the background, the visitor may view the result as arbitrary magic, draining the interactive experience of its theatrical momentum. The design must deliberately reveal the mechanism of interpretation, placing the user in the loop as an active collaborator and editor.

One Full "Show the Mechanism" Interaction

1. Input: The user types the sentence, "The future is a closed loop," and submits it to the machine.

2. Processing: The interface briefly displays a parsing animation, intentionally revealing the machine's labor as it evaluates the semantics.

3. Proposal: The 3D scene renders. A side panel slides out, displaying the LLM's semantic explanation in plain text: "The machine interprets this as highly enclosed and permanent, with low branching potential."

4. Exposed Controls: Below the text explanation, a series of sliders (representing Evaluation, Potency, Activity) and geometric dials (Enclosure, Erosion, Complexity) become visible. These are pre-set to the LLM's chosen values (e.g., Enclosure \= 1.0, Branching \= 0.0).

5. User Revision (The Edit Button): The user disagrees with the machine's interpretation, clicking and dragging the Enclosure slider from 1.0 down to 0.0. The 3D scene responds instantly, utilizing requestAnimationFrame updates to recalculate the InstancedMesh matrix and instantly disperse the geometry19. The visitor has successfully exerted their editorial power over the doctrine.

Explicit Unrecognized/Ambiguous-Input States

If a user submits keyboard smash (e.g., "asdfghjkl") or highly adversarial ambiguity, the LLM will struggle to map the input to the Osgood EPA dimensions. Instead of defaulting to a randomized, hallucinated visual state, the LLM prompt must instruct the model to populate the schema with a specific uncertainty boolean flag set to true.

• Visual State: Upon receiving this flag, the geometry enters a specific "Null/Becoming" state—a sparse, low-opacity, unmoving particle field with zero connectedness or structural hierarchy.
• Feedback: The UI explicitly states, "The doctrine cannot parse this meaning. Please clarify the sentence, or manually sculpt the world using the controls." This preserves the artistic premise of the machine having limits without breaking the software architecture.

7. Preservation of Authorship and Lineage

When a visitor alters a sentence, the resulting state must maintain a strict parent/version relationship. If user A creates a world, fine-tunes the sliders, and shares a remix link with user B, user B must see the exact identical geometry. Relying on the LLM to re-evaluate the sentence upon page load is unacceptable, as the stochastic nature of language models might interpret the identical sentence differently during a fresh generation.

To achieve exact replay and track the lineage of ideas, the system should adopt a provenance model based heavily on the W3C PROV-O ontology16. PROV-O is a standard for representing provenance—the documented history of how a piece of data was created, by whom, using what process, and from which inputs.

Provenance Model (ORIGINAL PROPOSAL)

• prov:Entity (The Artifact): The saved JSON schema containing the exact numerical parameters, the governing sentence, and the active archetype.
• prov:Activity (The Process): The act of interpretation (calling the LLM API) or the act of manual revision (the user physically dragging a slider).
• prov:Agent (The Author): Identifies whether the specific parameter change was authored by the human visitor or the machine model.

When a remix link is generated, the URL payload or backend database reference must point to the frozen prov:Entity (the specific JSON recipe), guaranteeing exact replay. The LLM is only invoked when a new sentence is submitted, generating a new prov:Entity linked to the parent entity via the prov:wasDerivedFrom relationship17. Private context, session IDs, and visual consent flags are strictly excluded from the remix export packet, treating the user text solely as portable data.

8. No-Code Semantic Evaluation and Reliability

If participants cannot reliably connect a visual transformation back to the edited sentence, the effect is mere spectacle. The system's visual-semantic mapping vocabulary must be empirically tested before launch.

Side-by-Side Evaluation Rubric

A no-code semantic evaluation can be conducted using matched before/after storyboards, entirely avoiding the collection of sensitive personal material from users. Raters from diverse linguistic backgrounds are shown a visual transformation (World A transitioning to World B) alongside three distinct sentence pairs:

1. The actual user input that generated the transformation (e.g., "One answer" \-\> "Many possibilities").

2. An adversarial pair with inverted semantic meaning.

3. A completely unrelated, randomized sentence pair.

Raters are tasked with matching the visual transformation to the correct sentence pair based purely on visual intuition.

Evaluating Reliability (Krippendorff's Alpha):

The agreement among human raters is measured using Krippendorff's alpha, a statistical measure of inter-rater reliability.

• [Figure omitted from source export]: The visual mapping is reliably interpretable and robust.
• [Figure omitted from source export]: The mapping is tentatively acceptable but requires refinement.
• [Figure omitted from source export]: The visual transformation is illegible and fails to convey the intended meaning45.

Question Answered: Specify what evidence would justify refining the mapping vocabulary, simplifying the model, or rejecting a proposed dimension.

If Krippendorff’s alpha falls below 0.667 for a specific visual dimension (for example, "Palette Saturation"), this constitutes concrete empirical evidence that the dimension is too culturally subjective, visually muddy, or misunderstood by the general public. In response, the design team must reject that specific proposed dimension, reverting control of that visual property to a static baseline. This simplifies the model and preserves the overall legibility of the remaining dimensions.

9. Minimum Viable Implementation Recommendation

To launch effectively while mitigating the severe risks of LLM latency, API cost, and hallucination, the Minimum Viable Implementation (MVI) must architecturally decouple the rendering engine from the AI completely.

1. Preserved Manual Mode: The core application operates as a standalone web-based procedural generator. The 15 visual dimensions are controlled entirely by HTML sliders. The Three.js canvas reacts in real-time to these slider changes without requiring a server connection.

2. Pre-Authored Examples: The 12 sentence-pair case designs are hardcoded as static JSON files on the server. Visitors can click through these transformations instantly, demonstrating the artistic premise of "The Indoctrination Machine" without a single live API call.

3. Optional AI Layer: The LLM is implemented as an optional serverless function that accepts a sentence and returns the JSON schema. If the API fails, times out, or the user is offline, the interface degrades gracefully. It defaults to the Manual Mode, displaying an error message that prompts the user to become the sole author of the doctrine.

This decoupled architecture guarantees that the rendering pipeline, accessibility protocols, and sharing mechanisms function flawlessly even if the AI components are entirely removed or suffer an outage.

10. Design Implications and Health/Accessibility Constraints

Any procedural 3D system featuring oscillation, erosion, and rapid geometric updates risks triggering photosensitivity if left unchecked. The system architecture must structurally prohibit flashes above 3 Hz to strictly comply with WCAG 2.2 Success Criterion 2.3.1 (Three Flashes or Below Threshold)31.

Consent Protocol:

As mandated by the client context, any potentially photosensitive patterned content must remain concealed behind an explicit site-visit opt-in. The visitor must actively choose between hidden/text-only, still artwork, or artwork plus motion before any WebGL canvas is initialized.

• To ensure the session does not infinitely persist without ongoing consent, the session flag must be stored in sessionStorage (which is naturally cleared when the browser tab closes) and feature a strict 30-minute interaction timeout.
• Motion permission does not equal autoplay permission. The JSON recipe output from the LLM or a remix link must never alter the user's local consent state.
• The requestAnimationFrame loop driving the motion must be gated by a global isMotionAllowed boolean. In "still artwork" mode, on-demand rendering is utilized—the scene is rendered once when parameters change, and the delta time multiplier is forced to zero, halting all continuous InstancedMesh updates and ShaderMaterial time uniforms immediately to save battery and prevent unwanted animation43.

11. Unresolved Questions and Five Prioritized Next Decisions

Unresolved Questions:

• How will the system handle multilingual input? While Osgood's Semantic Differential has cross-cultural validation, LLMs may exhibit bias in evaluating Potency and Activity in non-English languages.
• Can the WebXR implementation maintain the required 60FPS draw-call budget when rendering the "Chaos" archetype with high complexity counts?19.

Five Prioritized Next Decisions:

1. Solidify the JSON Schema: Finalize the exact variable names, floating-point bounds, and enumeration limits of the procedural recipe to lock in the API contract between the frontend developers and the prompt engineers.

2. Conduct Rubric Testing on the Archetypes: Render static storyboards of the Monolith, Chorus, and Unraveling states. Run a rapid Krippendorff’s alpha test with a sample of 20 users to ensure the baseline semantic intent is legible before committing resources to coding complex animations.

3. Establish Three.js Draw-Call Budgets: Set strict upper limits on the maximum InstancedMesh count and BufferGeometry complexity to ensure the experience achieves a stable framerate on mid-range mobile devices.

4. Implement the Opt-In Gate: Build the session storage logic for the hidden/still/motion choices and verify through testing that asynchronous API returns or deep-linked remix URLs cannot bypass this safety gate.

5. Refine the Prompting Logic: Develop the system prompt for the LLM to ensure it outputs raw, strictly validated JSON schemas with zero markdown wrapping or conversational filler, minimizing parsing errors and latency.

What this research would change in the experience

The following table delineates actionable changes for the design and development teams, distinguishing between changes justified by empirical evidence, features worth prototyping, and sensational claims that must be abandoned.

 

This is for informational purposes only. For medical advice or diagnosis, consult a professional.

Works cited

1. The Semantic Differential Scale | Appinio Blog, https://www.appinio.com/en/blog/market-research/semantic-differential-scale

2. Semantic differential \- Wikipedia, https://en.wikipedia.org/wiki/Semantic\_differential

3. Semantic Differential Scales: Theory, Construction, and Analysis, https://lensym.com/blog/semantic-differential-scale

4. A review of the limitations of Attention Restoration Theory and the, https://ojs.unito.it/index.php/visions/article/download/3323/3091/11293

5. What is Kaplan's Attention Restoration Theory (ART)?, https://positivepsychology.com/attention-restoration-theory/

6. The Magic of Nature \- Attention Restoration Theory and Fractals, https://yuvalbloch.com/blog/art/

7. Learning Shape Part Segmentation from Language Reference Games, https://arxiv.org/html/2112.06390v2

8. ShapeTalk: A Language Dataset and Framework for 3D Shape Edits, https://openaccess.thecvf.com/content/CVPR2023/papers/Achlioptas\_ShapeTalk\_A\_Language\_Dataset\_and\_Framework\_for\_3D\_Shape\_Edits\_CVPR\_2023\_paper.pdf

9. Language-Assisted 3D Shape Edits and Deformations \- ChangeIt3D, https://changeit3d.github.io/materials/changeIt3D\_main\_paper.pdf

10. 3D-GPT: Procedural 3D Modeling with Large Language Models, https://www.researchgate.net/publication/394944727\_3D-GPT\_Procedural\_3D\_Modeling\_with\_Large\_Language\_Models

11. LL3M: Large Language 3D Modelers \- arXiv, https://arxiv.org/html/2508.08228v1

12. Code-as-Room: Generating 3D Rooms from Top-Down View ... \- arXiv, https://arxiv.org/html/2605.18451v1

13. Dynamic Typography: Bringing Text to Life via Video Diffusion Prior, https://arxiv.org/abs/2404.11614

14. Dynamic Typography: Bringing Text to Life via Video Diffusion Prior, https://www.alphaxiv.org/abs/2404.11614

15. Dynamic Typography: Bringing Text to Life via Video Diffusion Prior, https://arxiv.org/html/2404.11614v2

16. PROV-O \- Open Neuroscience Graph, https://openneuroscience.org/Standards/PROV-O

17. PROV-O: The PROV Ontology \- W3C, https://www.w3.org/TR/prov-o/

18. 5\. Provenance information \- FAIR Cookbook, https://faircookbook.elixir-europe.org/content/recipes/reusability/provenance.html

19. InstancedMesh – three.js docs, https://threejs.org/docs/pages/InstancedMesh.html

20. WebGL 3D Facts and Information \- Three.js Resources, https://threejsresources.com/facts

21. CNS-Edit: 3D Shape Editing via Coupled Neural Shape Optimization, https://arxiv.org/html/2402.02313v1

22. Fine-Grained 3D Shape Editing via Primitive-Based Abstractions, https://arxiv.org/html/2604.23774v1

23. DesignAgent3D: Interactive 3D Scene Editing via Designer ... \- arXiv, https://arxiv.org/html/2608.21438v1

24. Panos Achlioptas \- alphaXiv, https://www.alphaxiv.org/@panos-achlioptas

25. ShapeTalk: A Language Dataset and Framework for 3D Shape Edits, https://cvpr.thecvf.com/virtual/2023/poster/21192

26. (PDF) Working with Interface Metaphors \- ResearchGate, https://www.researchgate.net/publication/300933278\_Working\_with\_Interface\_Metaphors

27. Conceptual Metaphor design pattern, https://ui-patterns.com/patterns/Conceptual-metaphor

28. Metaphoraction: Support Gesture-based Interaction Design with, https://zhdsun.github.io/papers/sun\_tochi22\_mim.pdf

29. Perceptual fluency and eye movements when viewing urban ... \- PMC, https://pmc.ncbi.nlm.nih.gov/articles/PMC12267524/

30. Full article: Attention Restoration Theory: A systematic review of the, https://www.tandfonline.com/doi/full/10.1080/10937404.2016.1196155

31. Success Criterion 2.3.1 \- Three Flashes or Below Threshold \- Level A, https://appt.org/en/guidelines/wcag/success-criterion-2-3-1

32. 37https://digital.library.unt.edu/ark:/67531/metadc332321/m2/1/high\_res\_d/1002782649-Magnavito.pdf

33. Psychological Perspectives on Aesthetic Experience \- IntechOpen, https://www.intechopen.com/chapters/1220466

34. Measuring aesthetic emotions: A review of the literature and a new, https://pmc.ncbi.nlm.nih.gov/articles/PMC5459466/

35. Understanding Semantic Differential Scales | PDF \- Scribd, https://www.scribd.com/document/217905692/Semantic-differential2-docx

36. Performances of LLMs in Multimodal Metaphor Understanding, https://www.sciencepg.com/article/10.11648/j.ijll.20261402.12

37. Text2BIM: Generating Building Models Using a Large Language, https://ascelibrary.org/doi/10.1061/JCCEE5.CPENG-6386

38. How JSON schema works for LLM data \- Latitude.so, https://latitude.so/blog/how-json-schema-works-for-llm-data

39. Three.js Games: Examples and How to Build One \- Seele AI, https://www.seeles.ai/resources/blogs/three-js-games-examples-how-to-build

40. Uniform – three.js docs, https://threejs.org/docs/pages/Uniform.html

41. https://threejs.org/docs/pages/ShaderMaterial.html

42. Three.js Crash Course: Interactive 3D for the Web \- Karyfy, https://www.karyfy.com/courses/three-js

43. Angular Signals \+ Three.js: How We Hit 60fps Without Going Zoneless, https://dev.to/emmanueln07/angular-signals-threejs-how-we-hit-60fps-without-going-zoneless-23jg

44. PROV-O: The W3C Provenance Ontology \- CASRAI, https://casrai.org/dictionary/term/prov-o

45. Introduction to Krippendorff's Alpha: Inter-Annotator Data Reliability, https://encord.com/blog/interrater-reliability-krippendorffs-alpha/

46. Inter-Rater Text Agreement Calculator \- MetricGate, https://metricgate.com/docs/inter-rater-text-agreement/

47. Reliability of psychiatric diagnoses in the 21st century \- PMC \- NIH, https://pmc.ncbi.nlm.nih.gov/articles/PMC13462468/

48. Scaling performance \- React Three Fiber, https://r3f.docs.pmnd.rs/advanced/scaling-performance