Built on the distinction between surface similarity and structural recognition. Three layers of representation. One knowledge graph that learns from use.
Layer 01
Three-level embedding
Every source is represented at three levels: surface vocabulary, conceptual structure, and a vocabulary-neutral essence abstraction. The essence layer is where “quantum vacuum” and “Brahman” become retrievable together.
surface · conceptual · essence
Layer 02
Hybrid retrieval
BM25 sparse search, dense vector search, knowledge graph traversal, and a reranking pass — run in parallel and fused. The retrieval is not a keyword search. It is a structural query against everything you have ingested.
BM25 · dense · graph · rerank
Layer 03
Living knowledge graph
Auto-populated during ingest. Recognition clusters form as sources connect. Vocabulary bridges link terms across traditions. The graph compounds with every source added — it does not reset or summarise.
auto-populated · multi-hop · compounding
Layer 04
Confirmation as training signal
Every time a researcher confirms or refines a cross-tradition connection, that judgment enters a training dataset. Over time, these confirmations calibrate a domain-specific model that learns this field’s vocabulary from expert behaviour — not from a generic prompt.
confirmation · vocabulary map · dataset
Layer 05
Passive intelligence distillation
Four training datasets accumulate from real usage: query-relevance pairs, essence translation pairs, entity-relation examples, and vocabulary equivalences. When thresholds are reached, a domain-specific adapter is trained — a model that improves retrieval precision in a specific field.
Mistral 7B · QLoRA · four datasets
Layer 06
Grounded generation only
Generation is grounded entirely in what has been ingested. When the knowledge graph cannot support a claim, the system surfaces the gap rather than filling it with world knowledge. Every output is cited and traceable to a source.
grounded-or-silent · cited · traceable