Retention-Aware Match Tracing for Auditable Allocation of Rural Youth Talent to Village Posts under Generative AI-Assisted Profiling

Abstract

County programmes that recruit young people into rural industry and village governance increasingly pair generative-AI profiling of unstructured application materials with algorithmic person-post matching, yet the matchers in use optimise immediate placement, explain themselves post hoc if at all, and are silent about the outcome the whole talent chain is built around: whether the young person is still in the village two years later. This paper proposes Retention-Aware Match Tracing (RAMT), a matching engine that is auditable by construction. The scorer is additive over a pre-registered ledger of person-post features, each passed through a piecewise-linear spline, so every match ships an evidence ledger that reconstructs its score exactly, with zero residual rather than an attribution estimate. The score feeds a discrete-time retention hazard trained on logged administrative matching records with inverse-propensity weights, and offers are formed by capacity-constrained deferred acceptance, with low-margin matches abstained to human review. Because no county authority can release youth-level records, evaluation uses a simulated labour micro-market whose marginals are anchored to published national statistics and whose ground-truth retention process is deliberately outside the model family — regime-switching, threshold-bearing and interactive. Across 20 market replications, RAMT converts 28.1 of every 100 offers into matches still in place at 24 months, 4.3 times the administrative wage-rank rule and 88.2 per cent of the oracle ceiling, while its ledger passes counterfactual-flip audits that sampled Shapley trails on equally accurate black-box baselines fail. The principal warning is distributional: because village wages sit below graduate expectations, a retention-optimal matcher rationally deprioritises the most educated youth, assigning them at 0.65 times the rate of their less-educated peers — the engine quietly gives up on exactly the people the policy is trying to attract, and repairing this costs a small, quantifiable share of total retention.