What problem does it solve? Observed and null diagrams are sometimes compared after independently re-fitting PCA, scalers, or loadings, which silently invalidates persistent-homology comparisons and forces results under PROVISIONAL flags. This Skill runs the diagnostic checks that detect such frozen-loadings defects, representation drift, and misuse of visualization-only projections before they corrupt claim-bearing inference. ## Core Features & Use Cases - Coordinate-Frame Auditing: Verifies whether observed and null data share one frozen fitted frame or are re-fit per draw, flagging invalid comparisons. - Representation Status Classification: Distinguishes inference (claim-bearing) representations from visualization-only outputs such as UMAP and t-SNE projections. - Reproducibility and Language Checks: Confirms seeds, scalers, PCA, and loadings are recorded and pinned, and enforces associational-only phrasing for SHAP explanations. - Use Case: Before running a W2 distance comparison between observed and null PCA embeddings, run this diagnostic to confirm both were transformed with frozen loadings; if not, the result is flagged PROVISIONAL and routed to contract-first-tdd. ## Quick Start Ask the agent to run the representation diagnostics on the ngram_pca embedding before comparing observed and null persistent homology results.
