Preguntas frecuentes
Goal
Transform scientific papers into interactive 3D visualizations where every visual element traces back to its source text, enabling researchers and students to explore complex spatial concepts with verifiable accuracy.
Why
Scientific papers describe spatial and temporal phenomena in text. Readers must mentally reconstruct 3D processes from 2D descriptions. This cognitive burden limits comprehension and creates opportunities for misinterpretation.
**The core problem:** A sentence like *"cylindrical tubes undergo shape changes to a string of pearls"* requires readers to imagine something most have never seen.
**The solution:** Interactive 3D visualization with provenance—click any shape, read the exact sentence that justifies it.
Concepts that work
Provenance tracking - Links every visual to source text, creating trust
Emphasis scoring - Allocates visual complexity by scientific importance
Causality layout - LEFT → CENTER → RIGHT spatial logic matches how processes unfold
Trigger-response maps - Encodes cause-effect as data, enabling interactive exploration
Story mode - Sequential reveal makes complex processes digestible |
|Manifest-driven architecture - Separates content from presentation, enabling flexibility |
Domain ontology separation** | Same engine works across biology, physics, economics |
Challenges
Monolithic codebase - Difficult to maintain, debug, or extend
LLM extraction variability - Same paper produces different results each run
Scope creep -Built features (audit systems, 10+ states) that were never used
No error boundaries - One bad input crashes the entire application
Misleading terminology - Called shader animations "physics simulation"
Missing infrastructure - No tests, no documentation, no design system
Next steps
Architecture
Modular file structure with single-responsibility components
Manifest-driven rendering (no hardcoded visuals)
Explicit dependency passing (no global state)
Error boundaries at every pipeline stage
Extraction
Deterministic baseline extraction (reliable without LLM)
Optional LLM enhancement layer
Human review interface for production use
Documented unit conversion formulas
Quality
Automated testing (unit, integration, visual regression)
JSON Schema validation for all manifests
Performance budgets (frame rate, load time)
Accessibility compliance (keyboard navigation, screen readers)
Documentation
Architecture decision records
Manifest authoring guide
API documentation per module
Domain ontology expansion guides
User Experience
Coherent visual design system
Responsive layout for multiple devices
Export capabilities (static HTML, video, images)
Shareable URLs with state
MVP
Requirements
Load manifest → render 3D scene
Click entity → show source sentence
Causality layout (left-to-right flow)
Story mode (sequential reveal)
PDF panel (side-by-side reference)
Deferred
Automatic LLM extraction
Real physics simulation
Advanced interaction modes
Collaboration features
Mobile optimization
Preguntas frecuentes
Tools used
3D Rendering |Three.js
Shaders | GLSL (WebGL 2)
PDF Display | PDF.js
Build System | Vite (recommended)
Language | TypeScript (recommended)
Testing | Vitest (recommended)
Why
Scientific papers describe spatial and temporal phenomena in text. Readers must mentally reconstruct 3D processes from 2D descriptions. This cognitive burden limits comprehension and creates opportunities for misinterpretation.
**The core problem:** A sentence like *"cylindrical tubes undergo shape changes to a string of pearls"* requires readers to imagine something most have never seen.
**The solution:** Interactive 3D visualization with provenance—click any shape, read the exact sentence that justifies it.
Metrics
Time to first render | < 2 seconds
Provenance coverage | 100% of entities
Frame rate | 60 fps desktop, 30 fps mobile
Manifest validation | 100% pass schema