Classify
deterministic intent rules + local neural intent model + local neural topic classifier + difficulty classifier
Cortex V6 is not presented as a mysterious local LLM. It is an owned intelligence core built from routing retrieval graph reasoning calibration memory planning evaluation and human review. Each subsystem has a specific job and failure boundary.
deterministic intent rules + local neural intent model + local neural topic classifier + difficulty classifier
BM25 and TF-IDF lexical evidence plus a local dense semantic hash encoder
knowledge graph expansion brings prerequisites contrasts and related concepts into candidate retrieval
lexical semantic graph scope exact-match and priority signals are combined into a hybrid score
confidence uses evidence agreement source coverage and top-result margin instead of copying a raw similarity score
Cortex selects a visible response plan based on intent topic difficulty memory availability and confidence
grounded extractive/template response with numbered evidence references and response-policy selection
answer quality checks grounding citation coverage confidence and uncertainty handling
signed-in learners can store visible deletable technical goals preferences and weak-topic memories
feedback updates response-policy Q-values while reviewed human teaching updates factual knowledge separately
The dense semantic encoder is locally implemented feature hashing. It is useful for semantic-ish retrieval but it is not equivalent to a pretrained embedding model.
Cortex exposes a compact response plan and evidence checks rather than claiming to reveal private internal reasoning.
Graph edges and human corrections influence retrieval only after scoring and review. Confidence is a calibration signal not a guarantee.