Mahalanobis Whitening

Alternative interpretation of IID mass-mean probing as projection onto θ_mm after Mahalanobis whitening

Neighborhood — ranked by edge-count

framework

Mass-Mean Probing
implements
Introduced in this paper: an optimization-free probing technique using difference-in-means direction with optional covariance correction

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

Embedding Whiteningconcept0.736
Preprocessing step using dev-set covariance to standardize span embeddings before computing S
whitening and z-scoring proceduremethod0.722
Calibration protocol: whiten embeddings on dev pool, z-score ρd and dr per layer.
whitening and z-scoring protocolmethod0.715
Standardization of ρd, dr, and log k on dev set for computing S.
Whitening of span embeddingsmethod0.708
Preprocessing step that uses dev-set covariance to standardize embedding scales before computing ρd and dr.
Dev-Set Whitening and Z-Scoring Protocolmethod0.688
Preprocessing pipeline for standardizing ρd, dr, and S across layers/models using dev-set covariance
Pigment-based paint mixingmethod0.680
Using pure pigments mixed in lime, cement, or other bases rather than tinted white-base commercial paints, allowing saturated, adjustable colors.
Gouache Paintingmethod0.667
Opaque watercolor method used for Veronica's blue chair, the investment banker's house interior, and color mockups.
Carewashingconcept0.661
Risk of companies labeling AI as mindful or compassionate for branding without genuine introspective architecture