Abstract
<title>Abstract</title> <p> Machine learning (ML) and operations research (OR) are routinely combined in supply chain decision support, yet the interface between them is often left methodologically unguarded: point or quantile forecasts are passed to an optimizer as though they were deterministic quantities, the resulting service levels are asserted rather than measured, and the marginal value of the stochastic machinery over a naive deterministic plan is rarely quantified. This paper presents a reference architecture that closes three specific gaps at that interface, together with an open-source reference implementation. First, the forecasting layer produces <bold>distribution-free calibrated prediction intervals</bold> via Conformalized Quantile Regression (CQR), with a monotone rearrangement operator that eliminates quantile crossing and an optional adaptive conformal update for non-stationary regimes. Second, the optimization layer replaces the common "solve deterministically on an inflated upper-quantile demand" heuristic with a <bold>Sample Average Approximation (SAA) procedure whose realized service level is empirically measured</bold> by Monte Carlo evaluation of the fixed route plan, including an explicit two-stage recourse cost model and a capacity-building wrapper that re-solves and re-validates against an independently seeded scenario draw. Third, the evaluation layer computes the <bold>Value of the Stochastic Solution (VSS) and the Expected Value of Perfect Information (EVPI)</bold> under a controlled comparison in which the deterministic and stochastic plans compete at identical vehicle capacity, so that the measured value is attributable to information rather than to slack. The architecture is instrumented for reproducible research: rolling-origin (walk-forward) cross-validation, proper scoring rules (pinball loss, approximate CRPS), calibration diagnostics (empirical coverage, interval width), Diebold–Mariano and Friedman–Nemenyi significance testing, optional MLflow tracking, a 53-function test suite, and containerization. We report the empirical result that has been executed and archived: a five-fold rolling-origin backtest on a 16-customer, 400-day synthetic logistics network. The uncalibrated quantile gradient-boosting ensemble attains CRPS 4.855 ± 0.433 and empirical coverage of 0.775 ± 0.030 against a nominal 0.80 target, <bold>under-covering in four of five folds (mean gap −2.47 percentage points)</bold> while forecast quality degrades monotonically across successive origins (CRPS +16.1% from the first to the last fold). This is a negative result about the uncalibrated baseline, and it is precisely the empirical motivation for the conformal layer: nominal quantile levels from independently fitted gradient-boosting models are not calibrated coverage levels, and treating them as such silently understates stockout risk. We state explicitly which components are validated by executed numerical evidence and which are implemented with validation harnesses whose outputs are not yet archived, and we outline the decision-focused learning and distributionally robust extensions that would convert this engineering contribution into a methodological one. </p>