Learning that can be inspected

Aner development mode records bounded variables, object references, collection mutations, and actual neural training captures. Offline reports show topology, separate fit loss curves, and available gradients. VS Code inspection panels and the optional Jupyter kernel display observations after cells complete. Production uses the same checked computation with automatic recording disabled.

Streaming charts, source linked pause/step debugging, and variable editing remain future work. Each recorded observation identifies its execution and step; historical captures and stale views are labeled, and viewing them does not rewind execution.

The evaluation goal is whether this workspace helps people explain a model failure and trace its data path. Aner should be assessed on workflow clarity and evidence. No comparative study or fixed performance improvement is claimed.