Spatial Limen: Structural Inference Engine
An evidence-grounded structural inference (finding hidden overlaps across documents) system designed to surface hidden overlaps across papers and patents while preserving provenance (a traceable record of where each piece of evidence came from), contradiction, and reviewer control.
- Role
- AI Product Builder
- Product
- Structural inference and research-discovery system
- Stack
- Next.js, Three.js, Supabase/PostgreSQL with pgvector, academic and patent ingestion pipelines
- Focus
- Provenance, contradiction handling, reviewer trust, and alternative user experience (UX) for deep discovery
Challenge
Standard search and basic RAG systems (retrieval-augmented generation, which grounds answers in approved source documents) are good at retrieval, but weaker at discovery. They work best when you already know what to ask, or when the relevant documents already live near each other through language or citation patterns. In deep research, the more valuable signal often sits somewhere else: when different fields are solving similar structural problems without using the same terminology, without citing each other, or while actively contradicting each other. That is where standard search and summary-first AI products tend to flatten nuance instead of surfacing it.
Build Decisions
Spatial Limen was designed to help users explore those hidden overlaps in a more structured way. Core capabilities include:
- Evidence-grounded discovery: Surface cross-domain overlaps without reducing everything to a generic summary.
- Provenance-preserving structure: Keep source, extracted evidence, and claims traceable rather than collapsing them into one layer.
- Contradiction-aware reasoning: Preserve support, challenge, and disagreement as their own tracked elements of the system.
- Explorable visual interface: Use a 3D graph interface for navigation and pattern recognition when chat is not the right tool.
- 27K+
- research nodes
- Review
- refusal-first gates
- Corpus
- papers + patents
- Traceable
- provenance graph
Inside the engine
The preview build leads with refusal, not discovery: candidate collisions (possible cross-document matches) face adversarial review (deliberately trying to disprove each match), refusals stay on record, and every node stays traceable to its source. Click any frame to enlarge.
Outcome Evidence
The strongest part of Spatial Limen is not just the interface. It is the trust model behind it. The system is built around a structured evidence layer that separates:
- Source from extracted evidence.
- Derivation from independence.
- Contradiction from deletion.
That matters because research systems become much less trustworthy when repeated claims look like independent support, or when conflicting evidence gets flattened away. Spatial Limen is designed to preserve lineage and disagreement instead of hiding them behind a smooth answer. On the product side, that leads to a very different design: ingestion and collision pipelines instead of chat-only retrieval, trust labels and reviewer logic instead of silent confidence assumptions, and visual exploration plus source traceability instead of black-box summarization.
- Provenance, contradiction, and source independence are preserved as their own tracked elements.
- Visual graph navigation complements retrieval for pattern discovery beyond chat summaries.
- System behavior is tuned for reviewer trust over answer smoothness.