MODEL COMPARISON
One illustrative scenario, four different formulations
All four models receive the same 5-day surplus, pantry, driver, timing, and daily food draws.
OPTIMIZATION POLICYBakedBoston Gurobi MIP
Maximizes one acceptance-adjusted expected-impact objective across simultaneous drivers, balancing food, fairness, environment, pantry priority, and driver fit.
Objective: Maximize the normalized food, fairness, environmental, pantry-priority, and driver-fit score after weighting route contributions by modeled acceptance.
Driver selection: The driver chooses recommendation rank 1, the highest-scoring route in the conflict-free menu produced after the joint assignment.
COMPARISON POLICYA minimal volunteer-route adaptation of the 2018 periodic unpaired pickup-and-delivery model: protect service first, then minimize miles.
Objective: Maximize assigned food-ready pickups, then minimize total route distance.
Driver selection: The model assigns one route to a driver; that assigned route is recorded as the driver's selection.
Not used to choose its routes: soft requested-time and ZIP preferences, acceptance probability, pantry fairness and priority, food-distribution fraction, CO2e.
COMPARISON POLICYA minimal volunteer-route adaptation of the 2025 Total-Curb model: protect service first, then minimize total direct system CO2e.
Objective: Maximize assigned food-ready pickups, then minimize uncollected waste, selected-route residual waste, and transportation CO2e.
Driver selection: The model assigns one route to a driver; that assigned route is recorded as the driver's selection.
Not used to choose its routes: soft requested-time and ZIP preferences, acceptance probability, pantry fairness, priority, and coverage, avoided production, meal preparation and packaging emissions, unobserved driver familiarity.
COMPARISON POLICYA minimal adaptation of the 2021 SLSF-noZ model: optimize short personalized menus over seeded stochastic willingness scenarios, then assign among routes drivers are willing to fulfill.
Objective: Maximize expected completed food-ready pickups over 100 seeded willingness scenarios, then minimize expected route distance as a tie-break.
Driver selection: The driver signals willingness for any acceptable menu options; the platform's recourse solve assigns one willing route, which is recorded as the driver's selection.
Observed menu trace: 2.93 options per driver on average · 88 total menu options · 1 willing drivers without a final assignment.
Not used to choose its routes: food quantity and pantry distribution as objectives, pantry fairness, priority, and coverage, CO2e as an objective, fare, compensation, and wage assumptions, unobserved unhappy-driver penalty history.
How to interpret the comparison
BakedBoston allocation: The joint Gurobi solve maximizes one normalized acceptance-adjusted expected-impact score: expected pantry coverage (10), raw donation volume (10), raw donation evenness (10), ultimately saved food volume (10), saved-food evenness (10), historical pantry opportunity priority (10), net direct CO2 benefit (20), and driver fit (20). Route contributions and pantry food totals are weighted by the transparent modeled acceptance probability; there is no 99%-of-best filter.
A driver receives no recommendation only when there are fewer distinct feasible bakery pickups than simultaneous drivers or that driver has no time-feasible route.
For BakedBoston's route-choice menu, the simulated driver selects rank 1, the highest-scoring route. The Horner adaptation instead follows its source formulation: drivers signal willingness for menu options and the platform makes the final recourse assignment. For other direct-assignment models, the assigned route is recorded as the driver's selection.
Expected acceptance and likely-rejection measures are prediction-based diagnostics; they do not override the deterministic recorded selection used in this demonstration.
All models share the exact same seeded surplus, daily food and usability draws, pantry openings and distribution fractions, waste allocations, driver events, facility windows, and feasible route geometry. Each selector reads only the inputs represented in its formulation; every result is then evaluated through the same food, environmental, travel, and acceptance ledger.
Balanced Total Impact: This scenario-relative index is not used by any optimizer, is not an externally validated social-impact measure, and can change when the scenario or comparison set changes. The underlying physical and fairness metrics remain the auditable result.
Seed 2033 is an illustrative balanced-tradeoff walkthrough in which all four policies complete the same number of deliveries. It is not presented as proof of average performance; multi-seed sensitivity analysis is the appropriate general comparison.
9bakeries
9pantries
38actual driver requests
1–3drivers sampled per window