An AI-Assisted Framework for Prioritising Roadside Friction Investigations Under Heterogeneous Urban Traffic

Roadside-friction investigation requires operational evidence and accountable engineering review. This study analyses 292 valid speed-bin records and 11,478 vehicle-frequency counts from eleven sections in the Malabe–Welivita corridor, Sri Lanka. Frequency-weighted analysis gives 59.47% combined-friction Relative Speed Loss (RSL). We calculate conditional Monte Carlo intervals for simple random monitoring samples drawn from the fixed empirical speed-bin distributions. At 100 vehicles per section, combined-friction RSL has median 59.49% and 90th percentile 60.63%. The investigation-priority score exposes policy-weight sensitivity: combined friction ranks first in 17/21 named specifications, ranging from first to sixth, whereas pedestrian 11–15/min ranks first more frequently under random-simplex weights. A frozen, logged, bounded contract-compliance audit compares deterministic lookup with 180 outputs from a large language model (LLM) under two advisory configurations across 30 cases. All three conditions have zero observed material failures on the defined automated contract checks, creating a ceiling that does not demonstrate superiority or improved engineering suitability. Automated checks do not establish engineering suitability or replace independent human scoring, which remains future work. The framework supports traceable investigation while retaining professional engineering authority over implementation.