Dense-seeded local diffusion is the retrieval policy used in this paper, positioned in Related Work against PageRank-style propagation. From a query concept, a dense encoder seeds the initial weight on the prerequisite graph and the procedure diffuses locally along graph edges. Two design choices differentiate it from PageRank: (i) additive hop penalties applied per traversed edge, and (ii) max aggregation over reaching paths to combine multiple paths to the same node, rather than scoring by the stationary distribution of a random walk with damping/teleport. The result is a deterministic, seed-localized score that is structurally similar to local PageRank in being seed-anchored but is not a stationary random-walk quantity.