This repository studies internal representations, representation stability, latent shifts, and failure modes. It does not claim that complex models can be completely explained.
This repository is not a generic XAI demo.
The repository remains inactive until the reliable-spatiotemporal-forecasting MVP creates a concrete diagnostic question.
It activates only when the flagship MVP exposes a concrete reliability failure.
Prediction-only versus prediction-reconstruction models under distribution shift.
Prediction-reconstruction is a candidate intervention, not an assumed solution.
The repository must not start from a preconceived claim that reconstruction improves representations.
Representation analysis is used to test hypotheses, not decorate reports.
Diagnostic methods must be connected to falsifiable questions.
- scope initialization only
- no modules
- no notebooks
- no diagnostic implementation
- no claimed findings