NEXUS CEREBRI / Learning and research
Where could learning help?
Cerebri began with a wish to understand how neural networks learn. Building the planner makes a related question concrete: could learned suggestions help find useful plans while every option still has to pass explicit checks? This is a research direction; the current planner follows fixed rules.
Rust already produces deterministic ranking observations. Evaluation Episode Phase A defines synthetic record types and checks their wire format; it does not collect observations or replay runs. The ML crate defines metadata and interfaces, with no trained model in the planner. Software 0.2.0 remains unreleased.
Later experiments could study request interpretation, preferences and search guidance. Each would need reproducible comparisons with the deterministic baseline. Facts, required rules, permissions and execution would remain outside learned authority.
Four strata show implemented Foundation and Temporal Core, deterministic planner verification and later provider, repair, learning and neural work as a separate research horizon.
Text alternative
Four strata show implemented Foundation and Temporal Core, deterministic planner verification and later provider, repair, learning and neural work as a separate research horizon.