Cognitive Agents for Bridge Inspection Prioritization: An Empirical Test of Predictability and Transparency Using the National Bridge Inventory (NBI)

Abstract

Objectives: This study develops a cognitive agent for highway bridge inspection prioritization and tests against transparent and statistical baselines, how well sustained condition deterioration can be predicted from National Bridge Inventory (NBI) data, and what role auditable reasoning plays when predictive signal is limited.

Methods: Six consecutive annual NBI snapshots for Connecticut, covering 2020 through 2025, were assembled into a longitudinal panel of 3,389 bridges. Four prioritization methods were compared: the fixed 24-month calendar schedule, a transparent rule-based risk index, a cross-validated statistical learner, and a large language model (LLM) agent grounded in inspection standards. Ground truth was defined from observed future condition under three outcome definitions of increasing specificity.

Findings: Predictability depends strongly on outcome definition. A single-point decline in any component over four years is nearly unpredictable for every method. Deck-specific deterioration and crossing into poor condition carry recoverable signal. Run over the full inventory, the cognitive agent matched the transparent rule-based index, with per-bridge priority scores correlating at 0.85, and both interpretable methods outperformed the opaque statistical learner on the policy-relevant poor-condition outcome. An independent expert review by a licensed bridge inspector rated the agent's reasoning at 2.77 out of 3, with 98 percent of rationales judged good or excellent.

Novelty: The study provides a reproducible benchmark of how outcome definition governs the predictability of bridge deterioration from inventory data, and positions language model agents within a measured accuracy-and-transparency tradeoff rather than asserting predictive superiority.

Practical Applications: Agencies adopting risk-based inspection should define prioritization targets precisely, because vague deterioration targets are not predictable from inventory data while specific targets are. Transparent methods remain competitive for well-defined outcomes, and the auditable rationale of a cognitive agent supports the human oversight that deployment requires.