Measuring Meaning in Retrieval Basins is a pre-registered longitudinal protocol by Nobel Glas for studying whether a publication set changes AI-generated answers over time.
The study freezes a panel of direct, adjacent, and unrelated control queries. It measures retrieval-based systems separately from primarily parametric systems and compares the entities, claims, and citations in later answers with a pre-deposit baseline. Jensen–Shannon divergence supplies the principal numerical measure.
The protocol’s central prediction is selective deformation: direct questions should change more than adjacent questions, while unrelated controls should remain stable. Provenance retention and temporal monotonicity are secondary tests. Failed predictions are to be deposited rather than hidden.
The paper contains no results. It also contains a control-design problem: one proposed comparison depends on a synthetic but plausible academic identity with fabricated prior deposits. That element is incompatible with transparent provenance research and should be replaced by a labeled, consented, nondeceptive control.