{"id":"6f627130-3f30-4243-a66c-60237ffdc0d2","arxiv_id":"2506.01694","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A scenario-cluster matheuristic solves a two-stage distributionally robust cross-dock door design model, giving bounds within 2.4 to 9.7 percent of a lower bound and matching or improving on CPLEX and Gurobi on the tested instances.","lead":"This paper builds a two-stage distributionally robust optimization model for designing cross-dock doors under uncertainty, and proposes a matheuristic that splits the problem by scenario clusters to get lower and upper cost bounds. It reports that this matheuristic matches or beats direct use of CPLEX and Gurobi on three test instances, and adds a risk-averse variant that limits costly outsourcing in rare scenarios.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported optimality gaps may be invalid because the SCD subproblem lower bounds are not certified: subproblems have a 4-hour time limit, and Table 9's SCD bound is below the LP bound in most rows.","rationale":"The reader's weakest assumption concerns the ambiguity set: the finite, hand-built set P is not validated against any out-of-sample distribution, and no statistical guarantee is offered. I agree that this is a legitimate gap, and it remains a reason for caution. However, the paper's headline claim is a numerical bound claim: the proposal is said to deliver incumbents with optimality gaps of 2.37-9.71%. Even if the ambiguity set were ideal, the Table 9 GAPs would still need to be true optimality gaps, which requires certified lower bounds. The 4-hour subproblem time limit, the absence of subproblem optimality certificates, and the fact that the reported SCD lower bound falls below the LP root bound in most rows make this the most load-bearing concern: it is an internal correctness issue that can be settled by re-running subproblems, rather than an external modeling debate. Credit is due where the argument is sound: the RN lower-bound logic in (7)-(8) is a standard relaxation argument and would be valid with certified subproblem optima; the upper-bound fixing scheme in (9) is a feasible construction for valid incumbents. The SD formulation is more delicate, and the paper itself flags the possibility of infeasibility with respect to (12c). The reader's CONDITIONAL verdict remains appropriate, but the conditions should explicitly include certified subproblem solves and corrected GAP reporting.","tokens_in":35353,"tokens_out":14936,"duration_ms":150682,"concrete_test":"Re-solve every SCD subproblem (7) and (16) for I3-RN-6, I3-SD-6, I7-RN-3, I7-RN-6 and I7-SD-3 with MIP gap tolerance 0.00% and no 4-hour limit, or record the remaining optimality gap at 4 hours. Recompute the bounds (8) and (17) using the certified optimal values, then recompute the Table 9 GAPs. If any recomputed lower bound is below the tabulated z_H, or if any subproblem was non-optimal at 4 hours, the reported GAPs are not true optimality gaps. Also verify that each fixed-first-stage SD incumbent satisfies constraint (12c); if not, the SD upper bounds are infeasible.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.4 defines the lower bound z_RN in (8) as max{z_LP-RN, max_{p in P} sum_{c in C_p} w~_c z^c_SCD-RN}; Section 4.3 does the same for z_SD in (17). These expressions are guaranteed lower bounds only if each z^c_SCD is the optimal value of subproblem (7)/(16), or at least a valid lower bound on it. The experiment in Section 5.4 states that each submodel is solved with a 4-hour time limit, and no optimality gaps are reported for the submodels. If a subproblem is interrupted at an incumbent, its returned value is an upper bound on the subproblem optimum, so the sum in (8)/(17) is no longer a guaranteed lower bound. Table 9 makes the problem visible: the reported SCD lower bound z_H is below the LP root bound z_L in six of eight rows (e.g., I3-RN-3: 6456.43 vs 7984.59; I7-RN-3: 17418.44 vs 20723.10). Since the paper's own GAP definition uses z' = max{z_L, z_H}, the headline 'optimality GAP 2.37-9.71%' is effectively measured against the LP root relaxation, not against a verified SCD bound. The claimed SCD contribution to the lower bound is therefore unsubstantiated, and the quantitative headline is not established. The SD side is additionally affected by the note in Section 4.3 that the scheme cannot guarantee feasibility with respect to (12c).","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the stochastic cross-dock door design problem under distributional ambiguity. It formulates two-stage mixed-binary-quadratic DRO models in risk-neutral and stochastic-dominance risk-averse versions, proposes a scheme to generate a finite ambiguity set by perturbing the nominal distribution and filtering candidates by Wasserstein proximity, and develops a scenario-cluster-decomposition matheuristic to compute lower and upper bounds. The computational study on instances I1, I3, and I7 compares the proposed bounds with CPLEX and Gurobi, reporting optimality gaps between 2.37% and 9.71% and goodness ratios between 0.957 and 1.001.","tokens_in":35706,"tokens_out":6800,"duration_ms":65003,"significance":"The problem addressed is relevant and computationally hard, and the modeling framework is coherent: the two-stage DRO formulation, the SD risk-averse extension, and the SCD-based bounding scheme are sensible design choices. The computational study is extensive and compares against two state-of-the-art solvers. However, the load-bearing quantitative claims are not established by the reported experiments: the lower bounds in Eqs. (8) and (17) are not certified, the SD upper-bound scheme is admitted to be potentially infeasible with respect to constraint (12c), and the ambiguity-set construction is not validated out of sample. If the bounding scheme were properly certified and the feasibility issue resolved, the matheuristic would be a useful practical tool for this problem. The paper does not provide machine-checked proofs, code, or data, so the numerical claims rest entirely on the described experiments.","major_comments":[{"comment":"The claimed lower bounds are not certified by the reported experiments. Each z^c_SCD is retrieved after solving submodel (7)/(16) with a 4-hour time limit (Section 5.4), but no optimality gaps for the submodels are reported; if a submodel is interrupted at an incumbent, its value is an upper bound on the subproblem minimum, so the weighted sum in (8)/(17) is not a guaranteed lower bound on z*_RN or z*_SD. Table 9 makes the consequence visible: the reported z_H is below the LP root bound z_L in six of eight rows (e.g., I3-RN-3: 6456.43 vs 7984.59; I7-RN-3: 17418.44 vs 20723.10), so with z' = max{z_L, z_H} the headline gap is effectively measured against the LP relaxation, not against the SCD bound. The paper should either prove each subproblem is solved to optimality, report subproblem optimality gaps, or replace z^c_SCD by a verified lower bound (e.g., the subproblem LP relaxation value).","section":"Section 3.4, Eq. (8), and Section 4.3, Eq. (17)"},{"comment":"The paper states that the proposed scheme cannot even guarantee that the solution of model LIP-SD is feasible for SD constraint (12c). This directly affects the upper-bound claim for the SD versions in Table 9. When the first-stage variables from submodel (16) are fixed and model (13) is solved, a solution violating (12c) is not feasible for the SD problem, so its cost is not an upper bound on z*_SD. The rows I7-SD-3 and I7-SD-6 report GR_H = ∞ because Gurobi finds no feasible solution; yet the paper later (Section 5.4) asserts that the SD bounds s_1 and \\bar{s}_1 are satisfied for these instances. The manuscript must report the verified surplus values \\sum_{\\omega\\in\\Omega_p} w_\\omega \\hat{s}^{\\omega,b} - \\bar{s}^b for each reported SD incumbent, or clearly label the SD results as heuristic costs for potentially infeasible designs.","section":"Section 4.3, Note"},{"comment":"The distributionally robust claim rests on the unvalidated premise that the small finite ambiguity set P selected by Wasserstein proximity to the nominal distribution adequately represents the true unknown distribution. The radius θ, the number of candidate perturbations |P'_q|, the noise variance σ_ε, and the SD triplet (ι_1, s_1, \\bar{s}_1) are all modeler-driven, and the SD triplet is chosen per instance specifically to eliminate outsourcing in the black swan scenarios. The observed improvement of SD over RN is therefore partly constructed by the modeler rather than discovered from data. A concrete remedy is an out-of-sample evaluation: generate additional realizations from the nominal mechanism, evaluate the RN and SD designs on them, and report worst-case and average costs as θ and the SD triplet vary. Without such validation, the title claim of a distributionally robust design is not established.","section":"Section 2.3 and Sections 5.1-5.3"}],"minor_comments":[{"comment":"The text says the first-stage vector is retrieved as in Step 2(b), but the vector is actually retrieved in Step 2(a); the cross-reference should be corrected.","section":"Section 3.4, Step 2(b)"},{"comment":"The header repeats 'z_H t_H' twice; rename the lower-bound and upper-bound columns (e.g., z_H^LB and z_H^UB) to avoid ambiguity.","section":"Table 9"},{"comment":"The sentence 'Set P is the subset of candidate members {p}, such that l_p^ρ ≤ θ up to p' is incomplete; the maximum cardinality |P| should be stated explicitly.","section":"Section 2.3"},{"comment":"There are several typos, including 'KEYWORS' and 'overperformances'; the manuscript should be proofread.","section":"Abstract and Section 1"},{"comment":"The Lagrangean decomposition is presented in detail, but no computational results for it are reported; state clearly whether LD is used in the experiments.","section":"Appendix B"}],"recommendation":"major_revision","confidential_remarks":"The paper would benefit from a reproducibility statement and from making instance data and the SCS4B subroutine publicly available, since the numerical claims depend on the companion paper Escudero et al. (2024b)."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Fernando,\n\nThe thing to know: the paper's headline lower-bound claim doesn't hold up against its own Table 9. The lower bound expressions (8) and (17) are only valid if each cluster subproblem is solved to optimality or gives a certified lower bound. The experiments state each submodel has a 4-hour time limit and no optimality gaps are reported; an interrupted solve returns an incumbent, which is an upper bound on the subproblem optimum. In six of the eight rows the reported SCD bound z_H is below the LP root bound z_L (e.g., I3-RN-3: 6456.43 vs 7984.59). Since the gap uses z' = max{z_L, z_H}, the headline 2.37–9.71% optimality gaps are effectively measured against the LP relaxation, not against a verified SCD lower bound. The SCD contribution to lower bounding is unsubstantiated.\n\nThat said, there is genuine value here. It is the first DRO treatment of cross-dock door design with a concrete ambiguity set generation scheme based on cdf perturbations and Wasserstein-distance selection. The risk-averse SD variant is a sensible extension, and the authors are transparent about its feasibility caveat in Section 4.3. The min-max matheuristic for upper bounds — fixing first-stage variables from cluster solutions and re-solving — is valid and produces incumbents that compare favorably with Gurobi (goodness ratios 0.957–1.001). The modeling is coherent and the paper is readable.\n\nOther soft spots: the SD parameter triplets are chosen per instance to eliminate outsourcing in black swan scenarios, so the RN-vs-SD comparison is partly constructed rather than out-of-sample. The ambiguity set is hand-built with no validation that it adequately represents the unknown distribution. No code or data are provided. These are fixable, and they are standard concerns for applied DRO papers.\n\nWho should read it? Supply chain planners and OR researchers working on cross-docking under uncertainty. The upper-bound heuristic is practically useful. But the lower-bound story needs major revision: either certify the subproblem bounds (report gaps or solve them exactly), or retract the claim that SCD strengthens the LP lower bound.\n\nRecommendation: send it to peer review. The model and the upper-bound work deserve referee time, and the lower-bound issue can be fixed in revision. I would not cite the lower-bound claim as-is, but I would consider citing the ambiguity set generation recipe.","headline":"Useful DRO engineering for cross-dock door design, but the headline SCD lower-bound claim is not supported by the numbers because the SCD bound falls below the LP root bound in six of eight runs.","tokens_in":36238,"tokens_out":3832,"would_cite":false,"duration_ms":36563,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C15","90C11","90B06"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a scenario-cluster-decomposition matheuristic supplies tight lower and upper bounds for two-stage distributionally robust cross-dock door design, with 2.37–9.71% optimality gaps on every tested instance.","keywords":["distributionally robust optimization","cross-dock door design","two-stage stochastic programming","scenario cluster decomposition","min-max matheuristic","Wasserstein ambiguity set","stochastic dominance","risk-averse optimization"],"falsifier":"Generate a hold-out scenario set from the same process that produced the nominal distribution, solve the proposed DRO design, and compare realized second-stage costs with the worst-case cost predicted over the ambiguity set; if realized costs substantially exceed the predicted worst case, the selected finite ambiguity set is not representing the true distribution.","tokens_in":35147,"feed_emoji":"🚛","tokens_out":8972,"duration_ms":88148,"temperature":0.7,"pith_summary":"This paper claims that the stochastic cross-dock door design problem—choosing strip and stack doors and their nominal capacities before uncertain inbound and outbound commodity volumes are known—can be handled by a two-stage distributionally robust optimization model in which an ambiguity set of probability distributions replaces a single unknown distribution. The authors propose generating that ambiguity set by perturbing the cumulative distribution functions of a nominal scenario set and retaining members within a chosen Wasserstein radius. For the resulting risk-neutral and stochastic-dominance risk-averse models, a scenario-cluster-decomposition matheuristic supplies lower and upper bounds. On their test instances it produces feasible incumbent designs for every run, with optimality gaps of 2.37–9.71% and costs within 0.957–1.001 of the Gurobi incumbent where one exists.","feed_headline":"2.37–9.71% gaps: min-max matheuristic bounds cross-dock design","feed_subtitle":"Scenario-cluster bounds beat Cplex and Gurobi on large risk-averse instances, where direct solvers find no feasible design.","key_machinery":"The machinery is the scenario cluster decomposition (SCD) built on split-variable reformulations of LIP-RN and LIP-SD. Copy variables for first-stage decisions (and for the robust cost u, and for the selector γ in the stochastic-dominance version) are introduced per cluster, with circular splitting-variable constraints (SVC) that are then relaxed; the resulting cluster submodels are solved independently. A lower bound comes from the maximum of the root LP bound and the weighted sum of optimal cluster submodel values, and an upper bound comes from fixing the first-stage solution of each cluster, solving the restricted full models, and taking the minimum over clusters of the first-stage cost plus the worst expected second-stage cost over ambiguity members. The ambiguity set itself is generated by cdf perturbations of the nominal distribution under four candidate probability densities, with weights recomputed by likelihood and members filtered by the Wasserstein transportation distance.","core_discovery":"Cross-dock door design under uncertainty is cast as a two-stage distributionally robust mixed binary quadratic program: the first stage chooses strip and stack doors and their nominal capacities, and the second stage assigns inbound and outbound commodity flows to those doors, with a penalized outsourcing option, under each scenario of each member of a finite ambiguity set. The paper builds the ambiguity set by perturbing the cumulative distribution functions of the nominal scenario set for four candidate probability distributions, reweighting scenarios by likelihood, and retaining members whose Wasserstein distance to the nominal distribution is small. For the risk-neutral model LIP-RN and the stochastic-dominance risk-averse model LIP-SD, a scenario-cluster decomposition relaxes the copy-consistency constraints between clusters, solves independent cluster submodels to get a lower bound, and fixes first-stage solutions from those submodels to compute an upper bound by re-solving the full restricted models. The computational study on instances with 5, 10, and 20 scenarios reports incumbents for all runs, optimality gaps of 2.37–9.71%, and goodness ratios of the proposed cost to the Gurobi incumbent between 0.957 and 1.001; for the largest stochastic-dominance instances, neither Cplex nor Gurobi finds any feasible solution in the 12-hour limit.","pith_inferences":["The SVC-relaxation plus min-max fixing scheme is transferable: any two-stage DRO with binary first-stage design decisions and scenario-wise recourse could be bounded the same way, not just cross-dock design.","A natural test the paper does not run is out-of-sample validation: evaluate the DRO design on held-out scenarios from the same process that produced the nominal distribution and compare realized costs with the predicted worst case.","The perturbation set size (20 per distribution, 80 candidates) and the Wasserstein radius are modeler-driven; enlarging them would show how sensitive the bounds and the robust cost are to the ambiguity set's coverage."],"forward_implications":["On every tested run the matheuristic returns a feasible incumbent within hours, with a certified optimality gap between 2.37% and 9.71%.","Where Gurobi returns a feasible solution, the matheuristic cost lies between 0.957 and 1.001 of Gurobi's, so the heuristic design is cost-competitive with direct solver use.","For the largest I7 stochastic-dominance models, Cplex and Gurobi both fail to produce any feasible solution within 12 hours, while the matheuristic produces one.","The stochastic-dominance variant eliminates outsourcing in the low-weight high-cost black-swan scenarios at a small increase in the robust objective value, while satisfying the surplus bounds."],"supporting_citations":[{"why":"Supplies the CDDP-TS two-stage stochastic model, scenario clustering scheme, and the SCS4B matheuristic that the DRO models and bounds build on.","marker":"Escudero et al. (2024b)"},{"why":"Defines the Wasserstein/earth-mover distance used to measure proximity of candidate ambiguity-set members to the nominal distribution.","marker":"Kantorovich (1942)"},{"why":"Provides the data-driven Wasserstein-metric DRO paradigm that motivates selecting ambiguity sets by Wasserstein radius.","marker":"Esfahani and Kuhn (2015)"},{"why":"Gives the linearization inequalities used to replace bilinear terms in the stochastic-dominance model.","marker":"Fortet (1960)"},{"why":"Provides the envelopes/Fortet-style inequalities used to linearize the mixed binary quadratic products in LIP-SD.","marker":"McCormick (1976)"},{"why":"Introduces DRO with stochastic dominance, the risk-averse functional adapted in model LIP-SD.","marker":"Dentcheva and Ruszczyński (2010)"},{"why":"Presents the cluster Lagrangean decomposition and SVC framework that the scenario cluster decomposition relies on.","marker":"Escudero et al. (2016)"},{"why":"Establishes coherence and time consistency of the stochastic-dominance risk measure used in the paper.","marker":"Escudero and Monge (2018)"}],"fun_headline_variants":["Scenario-cluster matheuristic beats Cplex and Gurobi on large instances","Robust cross-dock design: 2.37–9.71% gaps, direct solvers fail on large","Min-max matheuristic: scenario-cluster bounds for robust door design","Distributionally robust two-stage cross-dock design: matheuristic wins","Large-scale cross-dock design: matheuristic only feasible solver"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the small finite ambiguity set selected by Wasserstein proximity to the nominal distribution contains or adequately represents the true unknown distribution, so that minimizing the worst cost over it is genuinely robust; the paper offers no out-of-sample check of that premise.","fun_headline_variants_meta":{"raw":{"variants":["Scenario-cluster matheuristic beats Cplex and Gurobi on large instances","Robust cross-dock design: 2.37–9.71% gaps, direct solvers fail on large","Min-max matheuristic: scenario-cluster bounds for robust door design","Distributionally robust two-stage cross-dock design: matheuristic wins","Large-scale cross-dock design: matheuristic only feasible solver"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001096,"raw_usage":{"total_tokens":4635,"prompt_tokens":1065,"completion_tokens":3570,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":681,"completion_tokens_details":{"reasoning_tokens":3463}},"tokens_in":681,"tokens_out":3570,"duration_ms":25099,"temperature":1.0,"reasoning_tokens":3463,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:35:31.698925+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate a hold-out scenario set from the same process that produced the nominal distribution, solve the proposed DRO design, and compare realized second-stage costs with the worst-case cost predicted over the ambiguity set; if realized costs substantially exceed the predicted worst case, the selected finite ambiguity set is not representing the true distribution.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Wasserstein/earth-mover distance used to measure proximity of candidate ambiguity-set members to the nominal distribution."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the linearization inequalities used to replace bilinear terms in the stochastic-dominance model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the envelopes/Fortet-style inequalities used to linearize the mixed binary quadratic products in LIP-SD."},{"cited_title":"& Ruszczy\\'nski, A","cited_arxiv_id":null,"evidence_quote":"Introduces DRO with stochastic dominance, the risk-averse functional adapted in model LIP-SD."},{"cited_title":"& Monge, J.F","cited_arxiv_id":null,"evidence_quote":"Establishes coherence and time consistency of the stochastic-dominance risk measure used in the paper."}],"review_version":1}