Reliable Criterion Retrieval for Sensor-Instrumented Road Infrastructure: Diagnosing and Correcting a Title-Framing...

Retrieval-augmented generation increasingly serves as the knowledge backbone for sensor-informed infrastructure decisions, where a field engineer must retrieve the clause stating a design criterion, not a document about the topic. In twenty years (2005-2024) of Korea Expressway Corporation design-practice guidelines (HWP/HWPX), we identify a framing bias in dense retrieval as follows: embedding models over-weight the topical aboutness of a section title relative to the criterion in its body. The corpus invites this failure as follows: 63.2% of sections carry plan-framed titles, yet 86.4% of them carry criterion-type language in the body. A controlled counterfactual (n = 80) holding the body fixed and rewriting only the title isolates the effect (Δcos = 0.030, d_z = 1.63, p = 1.1 × 10^-14), reproduces it on a second embedding family, and decomposes it into term-frequency, early-position, and title-framing components; the framing residual (d_z = 0.53) survives primacy controls. A length- and frequency-matched neutral-token control splits that residual further into a token-composition component that replicates on both embedding families and a plan-framing component that reaches significance on one (d_z = 0.46). The bias buries framing-prone criterion documents by tens to hundreds of ranks; standard remedies are partial. We propose Criterion-Aware Retrieval (CAR), which hypothesizes the sought criterion at query time; on the 21 low-overlap queries that the bias hits hardest it outperforms both BM25 and the weighted-RRF hybrid after Holm correction (MRR 0.271 vs. 0.048 and 0.120). A cross-encoder reranker ranks better still (0.376) at no language-model cost but cannot exceed the recall of the pool it reorders (0.714 against CAR's 0.857): the two address different failure modes, and widening the pool is what the framing bias calls for. A parsing-pathway comparison shows that the structured pathways measured are near-lossless while PDF loses table content, justifying our HWPX-derived ground truth.