Where the SAE changes what gets retrieved

For each method, every one of the 12,146 navtest scenes is used as a retrieval query with both its original backbone and its SAE version. Δgreen = green_sae − green_orig (green = how many of the top 10 neighbours — cosine similarity, other driving logs only — have a 4 s future within 2.0 m ADE of the query’s), signed from −10 to +10: positive means the SAE retrieves more green neighbours (SAE better), negative means worse. Queries are ranked by Δgreen descending — rank 1 is where the SAE helps most, the last rank is where it hurts most — ties broken by the signed (mean ADE_orig − mean ADE_sae) descending (positive = SAE better there too), then by token. 50 queries are sampled at evenly spaced ranks (0% = SAE helps most, 100% = SAE hurts most) for inspection below.