LLM‑driven multi‑agent systems (MAS) are increasingly applied to high‑stakes decisions, yet audits that focus only on final outcomes often miss where fairness risks emerge along the decision trajectory. We introduce SCOPED‑Hiring, a process‑aware fairness diagnosis pipeline for LLM‑based hiring MAS. SCOPED‑Hiring creates controlled resume variants, assembles role‑based hiring committees, logs over 311 K structured decision trajectories, and translates trajectory fields into quantitative fairness signals organized under six diagnostic lenses: final outcome, counterfactual, process, pathway, dynamic, and design effects. The results show that balanced hire rates can conceal hidden unfairness in multi‑agent trajectories: career gaps trigger suspicion, proxy cues shape qualification judgments, and identity cues lead to unequal investigation. Targeted repairs guided by these diagnoses cut the total layered burden by 72.3% while shifting the hire rate by only 1.86 pp, demonstrating that process diagnosis can steer effective fairness remediation.
Review