Learning that can be inspected

Aner currently produces offline reports from actual neural network training captures, including topology, loss, and available gradients. Persistent notebooks also retain native typed values between cells.

A coordinated variable inspector, live charts, pause/step debugging, and linked network views are proposed additions. A recorded observation must identify its run and step; viewing an old capture must never imply that program execution has rewound.

The evaluation goal is whether this combined workspace helps people explain a model failure and trace its data path. Existing notebook and IDE tools already provide individual inspection features, so Aner should be assessed on workflow clarity and evidence, not on claiming those ideas are new.