{"id":"602c5ac7-8718-41f4-84c6-886f96b2cdfe","arxiv_id":"2504.18749","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Hybrid CP and RKO solvers combined with local search improve expected next-day delivery coverage by +20 to +50 basis points over a greedy baseline on a single real-world middle-mile network instance.","lead":"This paper applies constraint programming and random-key optimization to schedule truck departure times in an e-commerce middle-mile network, aiming to increase next-day delivery coverage. A hybrid version that feeds either solver's solution into a custom local search beat a production baseline by 20 to 50 basis points on one real-world instance, but standalone versions of both new solvers were less competitive.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"All reported bps gains are computed from an undisclosed black-box evaluator whose per-path aggregation appears to conflict with the stated DS-level overlap treatment; without a validated or reproducible objective, the hybrid-vs-baseline ranking is ungrounded.","rationale":"The reader's weakest assumption and my load-bearing concern align: the entire empirical comparison is mediated by an undisclosed black-box evaluator. I found no reason to move the verdict away from CONDITIONAL, but I did not find grounds to upgrade it either. The paper presents a plausible engineering study with a clearly described CP model and a reasonable RKO decoder, and the hybrid improvements could be real in the authors' industrial setting. However, the quantitative claim is not independently verifiable as reported. My additional point is that the text itself contains a possible internal inconsistency: Section 3.2 argues that overlapping SW coverage must be handled at the DS level to avoid double counting, while Eq. (3) sums over paths. This ambiguity matters because the reported basis-point gains are small and rounded to 10 bps, so even a modest evaluation error could flip the ranking. The proposed concrete test, replacing the unreleased oracle with a transparent held-out evaluator and comparing per-path versus DS-level aggregation, would settle whether the central claim survives without access to the proprietary black box. Since the reader already assigned CONDITIONAL, my stress-test reinforces that verdict rather than changing it.","tokens_in":17442,"tokens_out":5307,"duration_ms":59744,"concrete_test":"Rebuild the objective from first principles and rerun the final solver comparison. Concretely: (1) Specify a transparent evaluator, e.g., a deterministic simulation that computes realized next-day coverage from the TDT plan, leg transit times, node SLAs, DS cutoffs, and a held-out historical demand set; (2) train f(x) only on an earlier period and evaluate all final plans (baseline, RKO-hybrid, CP-hybrid) on a later hold-out period; (3) recompute the 1D KPI deltas of Tables 1 and 2. If the hybrid RKO and CP gains do not persist under this evaluator, the central claim is an artifact of the undisclosed f(x). As an internal cross-check, recompute F_black_box both as Eq.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim, that hybrid RKO and hybrid CP beat the baseline by +50 and +20 bps in 1D coverage (Table 2), rests entirely on F_black_box as defined by Eqs. (2)-(3). The core function f(x) is explicitly described as out of scope ('The specific implementation details for this black-box function f (x) are out of the scope of this paper', Section 2). This is not merely a reproducibility inconvenience: every KPI delta in Tables 1 and 2 is a difference of F_black_box values, so the direction and magnitude of every reported improvement is determined by a function the reader cannot inspect, reimplement, or validate against realized delivery outcomes. A second, internal concern sharpens this. Section 3.2 (Evaluate) states that overlapping SW-to-DS coverage must be handled at the DS level so that shared demand is not double-counted, yet Eq. (3) aggregates by summing f_black_box over all paths l. If f(x) is a per-path expected-coverage function, this per-path sum double-counts demand served through multiple warehouses; if f(x) is instead a marginal contribution, that is not defined anywhere. The normalization constant 'representing the maximum possible coverage of the network' (Section 4.1) is likewise undisclosed. The +50/+20 bps margins are small relative to the rounding precision of 10 bps and are selected as best-of-runs without repeated trials, so the result could shift with a defensible correction to the evaluator. In short, the paper's load-bearing evidence is an unverifiable oracle with a plausible aggregation inconsistency, and the headline comparison cannot be independently confirmed as stated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses the optimization of truck departure times in an e-commerce middle-mile network to maximize expected next-day delivery coverage. It formulates a constrained optimization problem with shift, capacity, labor-efficiency, and dispatch-spacing constraints, and an objective evaluated by an external black-box demand model. The authors develop a constraint programming (CP) model with a proxy package-speed objective and a random-key optimizer (RKO) with a decoder/propagator/violation-check architecture, and then test hybrid versions in which RKO/CP solutions seed a baseline greedy-plus-local-search solver. On a single EU-scale instance with 90 SWs, 34 SCs, and 242 DSs, the standalone solvers mostly trail the baseline in 1D coverage, while the hybrids report improvements of +40/+70 bps (single-wave) and +50/+20 bps (multi-wave) in 1D coverage, at the cost of 2D coverage. The RKO-hybrid requires days of warm-started compute versus hours for CP-hybrid.","tokens_in":17687,"tokens_out":5797,"duration_ms":56806,"significance":"If the reported gains are real, the work demonstrates that hybrid warm-starting of a domain-specific local search with CP/RKO solutions can improve next-day coverage at industrial scale, with practical implications. The paper also contributes a CP formulation with rolling capacity and labor constraints, a new RKO decoder design, and a careful treatment of multi-wave SW-SC legs. However, the evidence is entirely mediated by an undisclosed black-box evaluator, reported as single best-of-runs with no variance estimates, so the significance is conditional on the evaluator's fidelity and on the reproducibility of the 10-bps-scale deltas.","major_comments":[{"comment":"The aggregation rule is internally ambiguous and load-bearing. The text states that overlapping SW-to-DS coverage must be accounted for at the DS level to avoid double-counting shared demand, but Eq. (3) sums f_black_box(x,ℓ,Π̄) over all paths ℓ. If f_black_box is a per-path expected-coverage function, the sum double-counts demand that is reachable from multiple supply warehouses; if f_black_box is meant to be a marginal or DS-level contribution, that interpretation is never defined. Because every KPI delta in Tables 1 and 2 is a difference of F_black_box values, this ambiguity directly affects the reported +50/+20 bps improvements and must be resolved.","section":"Section 3.2 (Evaluate) vs. Eq. (3)"},{"comment":"The central objective is an undisclosed black-box function. The paper states that \"The specific implementation details for this black-box function f (x) are out of the scope of this paper,\" yet F_black_box is the sole objective used to score all solvers and all hybrid-vs-baseline deltas. Without at least a specification of f, a validation against realized delivery outcomes, or a reproducible implementation, a reader cannot determine whether the reported basis-point gains reflect real coverage improvements or artifacts of the estimator. This is a missing-support issue for the paper's central claim, not a minor reproducibility nicety.","section":"Section 2, Eq. (2)"},{"comment":"The results are single best-of-runs selected without repeated trials, and the deltas are close to the reported rounding precision (rounding to the nearest 10 bps). The +50 bps (RKO hybrid) and +20 bps (CP hybrid) improvements over baseline are within the range of a single rounding unit or one run's noise under the undisclosed evaluator. The paper should report means, standard errors, or best-of-k distributions over independent runs, and verify that the ranking is stable under changes to the penalty weight P and RKO hyperparameters.","section":"Section 4.2, Tables 1 and 2"},{"comment":"The comparison is run under asymmetric computational budgets. The RKO-hybrid result in Table 2 is \"derived from a warm start run, and so in aggregate it required multiple days of runtime,\" whereas the CP-hybrid is budgeted to a few hours; the baseline's runtime is not stated. Since the central comparison is about solver merit, the paper needs either a runtime-matched comparison or a plot of solution quality versus time; otherwise the +50 bps RKO result may simply reflect more compute.","section":"Section 4.1/4.2"}],"minor_comments":[{"comment":"The set notation \"w∈{1,..., W(i, j)]}\" uses a closing square bracket; it should be \"w∈{1,...,W(i,j)}\".","section":"Section 3.2, Random key encoding"},{"comment":"The word \"formualting\" should be \"formulating\".","section":"Section 3.1, Implementation details"},{"comment":"The word \"di ffs\" should be \"diffs\".","section":"Table 1 caption"},{"comment":"The CP approach is called \"exact\" in the abstract, but because it optimizes a proxy objective rather than F_black_box, the term is potentially misleading; consider calling it \"complete\" with respect to the CP model.","section":"Abstract and Section 3.1"},{"comment":"The text references \"3.2 and 3.2\" for Evaluate and ViolationCheck; these cross-references should be disambiguated.","section":"Algorithm 2 and Section 3.2"},{"comment":"The statement that small bps improvements are worth multiple millions of dollars is not sourced; a reference or sensitivity analysis would help.","section":"Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is an applied industry case study, and the undisclosed evaluator is the main risk. I would ask the authors to supply either the evaluator description or code in a supplement, plus repeated-run statistics; if that is not possible, the strength of the central claim should be substantially reduced in the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe one thing to know: the headline numbers—+50 bps for RKO-hybrid and +20 bps for CP-hybrid—come from an undisclosed black-box evaluator, and the paper's own Eq. (3) looks like it double-counts overlapping demand. Treat those gains as suggestive, not established.\n\nWhat is genuinely new is the modeling scope: sort centers, rolling time-window capacity, labor efficiency, dispatch spacing, and multi-wave legs on top of the earlier middle-mile work. The CP-SAT formulation is clearly written, with a stated proxy objective and explicit constraints. The RKO decoder, mapping random keys into IB/OB-feasible TDT bins, is a reasonable design, and the hybrid warm-start-plus-local-search protocol is a sensible way to combine the two.\n\nNow the soft spots. The black-box evaluator f(x) is explicitly out of scope; for an industrial paper that is acceptable, but here every KPI delta is a difference of F_black_box values. Without details or validation, the reader cannot reimplement or sanity-check the ranking. More concretely, Section 3.2 says demand must be aggregated at the DS level to avoid double-counting shared demand, yet Eq. (3) sums per-path f_black_box over all paths. If f is per-path, that double-counts; if it is a marginal contribution, that is never defined. The normalization constant for the KPIs is also undisclosed. The reported margins are best-of-runs, rounded to 10 bps, from a single instance, with RKO given days of compute while CP was budgeted to hours. Standalone RKO and CP underperform the baseline, so the hybrid local search is doing most of the work; that tempers the abstract's 'both approaches have merit' claim.\n\nWho gets value: anyone working on middle-mile logistics or applying CP/RKO to black-box objectives. The modeling extensions and decoder design are worth a look; the empirical claim needs stronger support.\n\nI'd send it to peer review, but with a clear request: disclose or validate the evaluator, fix the aggregation inconsistency, and report repeated trials with variance. If the authors can do that, the result may hold; as it stands, it's a single-run anecdote with an unverifiable objective.","headline":"A useful industrial case study whose headline gains rest on an undisclosed evaluator and an apparent aggregation inconsistency, so treat the bps improvements as plausible but unverified.","tokens_in":18341,"tokens_out":3233,"would_cite":false,"duration_ms":30566,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90B06","90C27","90C59"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper shows that hybrid solvers combining constraint programming or random-key optimization with local search can beat the incumbent greedy baseline on expected next-day delivery coverage.","keywords":["middle mile logistics","next-day delivery","truck departure time optimization","constraint programming","random-key optimization","black-box objective","local search hybridization","e-commerce logistics"],"falsifier":"Replace the black-box evaluator with actual next-day delivery outcomes on a holdout period: run the baseline, RKO-hybrid, and CP-hybrid plans through real operations, and check whether the hybrid plans still beat baseline by the reported margins; any reversal, or a margin within noise, would falsify the central claim. Repeating the comparison on an independent middle-mile network would test the single-instance generalisation.","tokens_in":17171,"feed_emoji":"📦","tokens_out":9211,"duration_ms":86069,"temperature":0.7,"pith_summary":"An e-commerce retailer's next-day delivery coverage depends on when trucks leave supply warehouses and sort centers in the middle-mile network, and this paper asks whether better truck departure times can be found than the incumbent greedy-plus-local-search solver. It reports yes, on a real EU-scale network with 2187 legs and 11,830 paths: hybrid versions of both a constraint-programming solver and a random-key optimizer beat the baseline in expected next-day coverage, by +40 to +70 basis points in the single-wave setting and by +20 to +50 basis points when sort-center legs carry two waves. The random-key hybrid posts the best multi-wave result (+50 bps) but takes days to run, while the constraint-programming hybrid gives +20 bps in hours. The practical stakes are large because a few basis points of next-day coverage at this volume are worth millions of dollars in faster deliveries, although the gains come at the expense of some two-day coverage.","feed_headline":"Hybrid solvers lift next-day delivery coverage by 50 basis points","feed_subtitle":"Random-key and constraint-programming hybrids beat a greedy baseline on a real logistics network.","key_machinery":"The load-bearing mechanism is the decoder-plus-fitness loop: in the random-key optimizer, a deterministic decoder maps a key vector $\\chi\\in[0,1)^N$ onto leg-specific truck-departure bins that respect inbound and outbound shift windows, then propagates those departures along each path, queries the black-box evaluator for expected coverage, and returns fitness $F_{\\text{black box}} - P\\cdot N_{\\text{viol}}$. The constraint-programming branch instead encodes the same physical constraints with integer variables and maximizes an analytically expressible \"package speed\" proxy, since the black box cannot be embedded in the search. The two are hybridized by using either solver's output to warm-start a local search, and a \"law-of-physics\" latest-feasible-departure calculation gives both branches a useful starting point.","core_discovery":"The central claim is that hybridization is what makes the solvers competitive: RKO or CP alone cannot beat the baseline on one-day coverage, but each wrapped in the baseline's local search routine can. The CP model maximizes a \"package speed\" proxy over integer truck-departure-time variables under shift, capacity, labor-efficiency, and multi-wave dispatch-spacing constraints. The RKO decoder maps random keys to feasible departure bins, propagates the plan through the network to derive path promises, scores it with the black-box coverage evaluator, and subtracts a penalty for constraint violations. In the multi-wave problem the RKO hybrid improves one-day coverage by +50 bps and the CP hybrid by +20 bps over baseline, while in the single-wave problem they improve by +40 bps and +70 bps respectively, in both cases at the cost of lower two-day coverage.","pith_inferences":["If the black-box coverage evaluator faithfully reflects real delivery outcomes, then even 20–50 basis points of next-day coverage is a measurable business win at e-commerce scale, and the hybrid design is the practical lesson; if the evaluator is not faithful, the ranking of solvers could change.","The single EU-scale instance leaves open whether the hybrids' edge survives on networks with different topology, volume mixes, or cutoff times; a second independent network would be the natural check.","Because the CP pruned domains were already used to restrict RKO decoding, a tight pipeline that warm-starts RKO from a full CP solution is a natural extension that could combine CP's speed with RKO's search.","A multi-objective variant that rewards next-day coverage while penalizing drops in two-day coverage would test whether the observed 2D loss is an artifact of the single-objective formulation or a structural tradeoff."],"forward_implications":["Re-optimizing truck departure times can raise next-day coverage without any change to warehouses, trucks, or delivery stations, purely by choosing when volume leaves each node.","Hybridization is the decisive design choice: the standalone RKO and CP solvers do not beat the baseline on one-day coverage, while both hybrids do.","The CP hybrid delivers most of the RKO hybrid's coverage gain at a fraction of the runtime, giving operators a compute-versus-quality tradeoff.","Adopting either hybrid shifts coverage from two-day toward next-day delivery, so the value of the gain must be weighed against the lost two-day promise.","Placing more truck departures than the baseline (the CP solver places the maximum on every leg) does not by itself improve coverage, confirming that the objective is non-linear."],"supporting_citations":[{"why":"It defines the earlier middle-mile TDT problem that this work extends with sort centers, labor efficiency, and multi-wave constraints.","marker":"[2]"},{"why":"It supplies the constraint-programming model-and-solve paradigm used by the CP approach.","marker":"[3]"},{"why":"It supplies the random-key optimizer architecture with decoder and search engine used by the RKO approach.","marker":"[4]"},{"why":"It provides the constraint-programming solver used to execute the CP model.","marker":"[12]"}],"fun_headline_variants":["Hybrid RKO and CP solvers lift next-day coverage by up to 50 bps","Hybrid solvers beat greedy baseline on next-day coverage in real network","Random-key and constraint hybrids improve next-day coverage vs baseline","Hybrid approach adds 50 basis points to next-day delivery coverage","Hybrid solvers outperform greedy in next-day coverage optimization"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison stands on the fidelity of the undisclosed black-box evaluator $f(x)$; if that learned coverage score does not track real next-day delivery outcomes, every reported basis-point gain is ungrounded, and the single EU-scale instance cannot compensate.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid RKO and CP solvers lift next-day coverage by up to 50 bps","Hybrid solvers beat greedy baseline on next-day coverage in real network","Random-key and constraint hybrids improve next-day coverage vs baseline","Hybrid approach adds 50 basis points to next-day delivery coverage","Hybrid solvers outperform greedy in next-day coverage optimization"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000424,"raw_usage":{"total_tokens":2215,"prompt_tokens":1028,"completion_tokens":1187,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":644,"completion_tokens_details":{"reasoning_tokens":1093}},"tokens_in":644,"tokens_out":1187,"duration_ms":10998,"temperature":1.0,"reasoning_tokens":1093,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:09:39.924771+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replace the black-box evaluator with actual next-day delivery outcomes on a holdout period: run the baseline, RKO-hybrid, and CP-hybrid plans through real operations, and check whether the hybrid plans still beat baseline by the reported margins; any reversal, or a margin within noise, would falsify the central claim. Repeating the comparison on an independent middle-mile network would test the single-instance generalisation.","supporting_citations":[{"cited_title":"Rossi, P","cited_arxiv_id":null,"evidence_quote":"It supplies the constraint-programming model-and-solve paradigm used by the CP approach."},{"cited_title":"Perron, V","cited_arxiv_id":null,"evidence_quote":"It provides the constraint-programming solver used to execute the CP model."}],"review_version":1}