Train
Owned labeled examples train local routing models and export versioned artifacts.
CortexLab V7
Local models vector search safe execution and bounded autonomous workflows.
V7 turns CortexLab from an intelligence pipeline into an engineering workbench. Models can be trained locally. Owned knowledge can be indexed. Code can run without touching the server. Agents can execute visible multi-step workflows with explicit stopping boundaries.
2packaged local models
146vector records
192Dsemantic vectors
71LSH buckets
V7 architecture
Four new capabilities with explicit boundaries.
Index
192D semantic vectors are grouped into locality-sensitive hash buckets then cosine-reranked.
Execute
Code runs in a disposable browser worker with blocked network APIs and a hard timeout.
Plan
Agents decompose research study debug comparison and content goals into bounded steps.
Retrieve
Each agent step uses Cortex hybrid retrieval and knowledge graph evidence.
Govern
Training runs agent steps memory graph edges and confidence telemetry remain inspectable.
Local model training
The built-in trainer is a compact multiclass linear model over hashed unigram and bigram features. It is useful for routing. It is not a frontier language model.
Vector index
Cortex uses local 192D feature vectors and LSH buckets to narrow candidates before cosine scoring. It can be rebuilt from owned CortexLab knowledge.
Bounded autonomy
Agents have explicit step limits and a fixed tool set. V7 deliberately does not give an autonomous loop unrestricted shell browser or external account access.