{"id":"179e1039-e817-491d-a089-509f5628340c","arxiv_id":"2509.06157","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"This paper defines meal kit box allocation as a MILP and reports that CBC solves synthetic instances up to 100,000 orders optimally in under two minutes.","lead":"This paper introduces a production planning problem for meal kit delivery, assigning customer orders to factories to keep daily recipe mixes stable. The authors report that an off-the-shelf MILP solver solves synthetic instances with up to 100,000 orders in under two minutes, beating two heuristic methods.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Model is under-specified: a_{i,j,t} is never linked to x in Eqs. (1)–(5), and Eq. (1) has a dangling site index in the denominator; optimality results are therefore not reproducible from the text.","rationale":"The reader's weakest assumption identifies exactly the same gap. I read the formulation carefully: constraints (2)–(5) assign orders to factories and enforce capacity and eligibility, but nothing computes recipe quantities a_{i,j,t} from x. The text's 'indirectly controlled' is not a constraint. As a result, the objective in Eq. (1) is not a function of the decision variables, so CBC's optimality claims and the WMAPE numbers cannot be checked from the manuscript. There is no formal verification or independent implementation. I do not think the underlying problem is fundamentally hard—it resembles an assignment/min-cost-flow model, and the fast exact results may be repairable—but as written the paper does not support its strongest claim. The reader's REJECT is appropriate; no change needed.","tokens_in":10707,"tokens_out":7075,"duration_ms":73081,"concrete_test":"Add the explicit linking constraint a_{i,j,t} = Σ_{o: i ∈ R_o} x_{o,j,t} (and the analogous constraint for day t−1) to the formulation, and fix the dangling denominator in Eq. (1), e.g., by using Σ_i a_{i,j,t} for each factory j. Then rerun the published GitHub code (pinned to a specific commit) on the 10,000-order benchmark of Table 6. If the reported WMAPE site = 0.054 and optimization time = 2.89 s are unchanged, the omission was cosmetic; if they change, the claimed optimality results rest on an unstated construction. Independently re-solve the repaired MILP with CBC and compare optimal objective values.","verdict_should_be":"UNCHANGED","load_bearing_attack":"To support the abstract's claim that CBC consistently achieves optimal solutions, the objective in Eq. (1) must be a well-defined function of the binary variables x. It is not. The quantities a_{i,j,t} and a_{i,j,t−1} appear in the objective, but no constraint or definition equation connects them to x_{o,j,t}. The informal sentence after Eq. (5) that they are 'indirectly controlled' is not a mathematical link. With only constraints (2)–(5), any feasible x determines only order-to-factory assignments; the objective is flat unless one adds something like a_{i,j,t} = Σ_{o: i ∈ recipes(o)} x_{o,j,t}. Additionally, Eq. (1)'s denominator Σ_i a_{i,j,t} keeps the index j free; as written the quotient is a vector, not a scalar, so 'minimize f(x)' is not well defined. Without these definitions, the WMAPE values in Table 6 and Figures 8–19, and the claim that CBC reaches a proven optimum, cannot be checked. A competent re-implementation could plausibly fix this with a linking constraint and by linearizing the absolute values using auxiliary variables, but the paper as submitted does not specify the model that was solved.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces the Box Allocation Problem (BAP) for meal kit delivery: assign orders to factories over a 15-day horizon while minimizing day-to-day recipe allocation variation, subject to capacity and recipe-eligibility constraints. It claims to formulate BAP as a MILP in Eqs. (1)-(5), compares the CBC solver against Tabu Search and Iterative Targeted Pairwise Swap on simulated instances up to 100,000 orders, and reports that CBC consistently reaches the WMAPE global lower bound in under two minutes, including under dynamic capacity and demand changes. The central computational claim is that CBC produces provably optimal allocations at commercial scale.","tokens_in":11035,"tokens_out":9612,"duration_ms":120108,"significance":"If the formulation and results were correct, the paper would be a useful applied contribution: it identifies a temporal allocation problem with a clear sustainability motivation, benchmarks an exact solver against heuristics, and proposes a lower-bound optimality check. The paper also makes its data and code available on GitHub, which is a strength. However, the mathematical formulation as written is not well posed: the objective in Eq. (1) is not a well-defined function of the decision variables, is not a MILP as claimed, and the stated lower-bound argument in Eq. (8) is invalid as written. These issues are load-bearing because the optimality claims, tables, and figures all depend on them. The significance of the paper therefore cannot be assessed until the model is correctly specified and the experiments are repeated or verified under that specification.","major_comments":[{"comment":"The objective is not a well-defined function of x. The notation lists a_{i,j,t} and a_{i,j,t-1}, but no constraint or definition links them to x_{o,j,t}. The sentence after Eq. (5) that these quantities are 'indirectly controlled' is not a mathematical relation. With only constraints (2)-(5), f(x) is constant with respect to the decision variables, so the reported optima in Table 6 and Figures 8-19 are undefined. The authors must add linking constraints, e.g., a_{i,j,t} = sum over orders o containing recipe i of x_{o,j,t}, and state how multi-unit recipe quantities within an order are counted.","section":"Section 2.2, Eq. (1)"},{"comment":"Even after adding a linking definition, Eq. (1) is not a MILP. It contains absolute values and a ratio, and the denominator retains the free index j while the numerator sums over j; as written the quotient is a vector indexed by j, not a scalar objective. The paper must specify a scalar objective (e.g., sum over factories of per-factory ratios, or a single ratio with a global denominator) and describe the standard linearization using auxiliary variables. This is necessary to substantiate the claim that CBC solves a MILP.","section":"Section 2.2, Eq. (1)"},{"comment":"The claimed lower-bound relation is not established. WMAPE site as defined in Eq. (1) has denominator sum_i a_{i,j,t} for a factory j, while Eq. (8) uses the global denominator sum_i a_{i,t} and labels the ratio WMAPE site. The triangle inequality only gives |sum_j u_j| <= sum_j |u_j|; it does not yield Eq. (8) when denominators differ. Consequently, the numerical check 'WMAPE site = WMAPE global' in Table 6 is not a valid optimality certificate. If the intended site metric is actually total absolute deviation divided by global volume, the terminology and equations must be corrected consistently.","section":"Section 2.2, Eqs. (6)-(8) and Table 3"},{"comment":"The experimental setup is incompletely specified. The model defines T = {-18,...,-3}, but the benchmark and scalability tests use 'two consecutive days LD12 and LD11'; it is unclear whether T is restricted to two days and what t and t-1 denote. The reported model size of 300,600 columns for 100,000 orders and 3 factories is also inconsistent with 100,000 x 3 x 2 = 600,000 (or 4.5 million over 15 days). The authors must specify the instance generator, factory capacities, initialization of previous-day allocations, and the exact model passed to CBC for each reported run.","section":"Section 4.2.1-4.2.2"}],"minor_comments":[{"comment":"The set T = {-18,-17,...,-3} contains 16 days, not 15 as stated throughout. Either exclude one endpoint or correct the count.","section":"Section 2.1 / Notation"},{"comment":"The quantities a_{i,t} and a_{i,t-1} are used in Eqs. (6)-(8) but are not defined relative to a_{i,j,t}. Add a_{i,t} = sum_j a_{i,j,t}.","section":"Section 2.2"},{"comment":"The reported improvement percentages are inconsistent with the rounded values: (0.074-0.054)/0.074 is approximately 27.0%, not 26.21% for ITPS. Recompute or state the exact unrounded values used.","section":"Table 6"},{"comment":"Reference [16] is cited as an 'Internal company document' and is used for the problem definition and Figures 2-4. This is not verifiable by readers. Please replace it with public sources or provide a detailed description in the paper.","section":"References / Figures"},{"comment":"The terms 'B&B' and 'CBC' are used interchangeably. Clarify that CBC is the solver implementing branch-and-bound, and use one consistent name for the exact method.","section":"Throughout"},{"comment":"No information is given about the number of heuristic runs or seed variability. Report means, standard deviations, or at least the number of repetitions for the heuristic results.","section":"Section 4.2.2"},{"comment":"The paper says data are on GitHub, but the reproducibility of the experiments would be greatly improved by including the exact CBC model files or a script that generates instances and solves them.","section":"Data Availability"}],"recommendation":"major_revision","confidential_remarks":"The formulation issues are so central that acceptance cannot be recommended now. However, they appear fixable in a revision: add the missing linking constraints, rewrite Eq. (1) as a linearizable scalar objective, correct the lower-bound argument, and rerun/verify the experiments under the corrected model. If the authors cannot supply the corrected model and reproducible runs, the paper should be rejected. Also, the reliance on internal company document [16] for the problem definition raises provenance concerns that should be checked by the editor."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a real-world optimization paper with a broken core formulation. The missing constraint linking a_{i,j,t} to x_{o,j,t} means Eq. (1) is not a function of the decision variables, so the reported optimality results are undefined. The dangling index in the denominator makes it worse. That said, the problem is genuine, the WMAPE global lower bound is a nice touch, and the fix is straightforward. It deserves a referee, not a desk reject.\n\nWhat's actually new: the BAP naming and the meal kit context, plus the temporal objective that penalizes day-to-day recipe changes. The comparison of CBC, TS, and ITPS is systematic, with plausible synthetic data. The triangle inequality bound is a legitimate optimality check, and the proxy-optimization argument is sensible.\n\nSoft spots: the missing definition is load-bearing. The 'indirectly controlled' sentence after Eq. (5) is hand-waving, not math. Eq. (1) has a dangling site index in the denominator, so the quotient is a vector. The novelty is limited: once repaired, the model is a standard transportation/assignment MILP with an L1 objective; the newness is in the domain framing and the temporal twist. The reproducibility gaps are secondary: synthetic data, single runs, no seeds or commit hash, but those are minor relative to the model gap.\n\nWho this is for: practitioners in meal kit logistics or multi-factory production planning, and researchers working on assignment problems with temporal objectives. They'll see the value once the model is properly specified.\n\nRecommendation: send to peer review with major revision requests, centered on writing down the linking constraint a_{i,j,t} = sum over orders of x_{o,j,t} for recipes in the order, fixing the denominator, and either linearizing the absolute values or stating how the MILP handles them. As written, the paper shouldn't be accepted, but it's a salvageable contribution worth a referee's time.","headline":"Interesting industrial application, but the MILP is under-specified as written: a_{i,j,t} is never linked to x, so the objective and the optimality claims are undefined; still, the problem is real and the fix is straightforward, so it deserves a referee.","tokens_in":11543,"tokens_out":2090,"would_cite":false,"duration_ms":27033,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C11","90B06","90C59"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper introduces the Box Allocation Problem (BAP) for daily assignment of meal-kit orders to factories and claims that an open-source MILP solver certifies optimal allocations at up to 100,000 orders in under two minutes.","keywords":["Box Allocation Problem","meal kit delivery","mixed-integer linear programming","CBC solver","tabu search","WMAPE","production planning","supply chain optimization"],"falsifier":"Run the released code on a two-factory, two-order instance where both orders contain recipe 1 and are eligible for both factories, and print the objective for both feasible assignments. If the objective value is identical for both assignments, the model has no decision-variable dependence, and the reported optimality is vacuous. The GitHub repository named in the data-availability statement is the place to look.","tokens_in":10581,"feed_emoji":"📦","tokens_out":5763,"duration_ms":66088,"temperature":0.7,"pith_summary":"This paper proposes a new optimization problem, the Box Allocation Problem, for assigning recipe-box orders to production factories day by day. The goal is operational stability: each factory's mix of recipes should change as little as possible from one day to the next, so ingredient forecasts stay accurate and food is not wasted. The authors model BAP as a mixed-integer linear program with capacity, eligibility, and single-assignment constraints, and benchmark the open-source CBC solver against Tabu Search and an iterative pairwise-swap heuristic. On simulated instances with up to 100,000 orders, they report that CBC reaches proven optimal solutions in under two minutes, including under changing factory capacities and changing customer orders. If correct, the paper's practical claim is that a free solver can certify optimal daily allocations at commercial meal-kit scale, without custom heuristics.","feed_headline":"Free solver nails meal-kit box allocation at 100,000 orders","feed_subtitle":"A mixed-integer model stabilizes daily factory recipes, shrinking forecast error and food waste in meal kit delivery.","key_machinery":"The load-bearing object is the MILP formulation in Eqs. (1)-(5): binary variables x_{o,j,t} decide which factory j receives order o on day t, with constraints (2) full capacity for all factories except the catch-all, (3) eligibility (an order goes only to a factory that can make every recipe in it), and (4) one factory per order. The objective is WMAPE site—the weighted mean absolute percentage error of recipe quantities per factory between consecutive days. The paper's computational engine is the CBC solver, whose primal heuristics (Feasibility Pump, Coefficient Diving) find the optimum at the root node and whose limited branch-and-bound certifies it. Because the objective's recipe quantiti","core_discovery":"On the paper's own terms, the central claim is that BAP is a genuine new problem—distinct from bin packing and vehicle routing because its objective is temporal—and that it can be solved exactly and quickly. The objective minimizes the site-level Weighted Mean Absolute Percentage Error (WMAPE), i.e. the sum over factories and recipes of absolute day-to-day changes in allocated recipe quantities, divided by the day's total recipe quantity. The MILP (Eqs. 1-5) controls the a_{i,j,t} quantities indirectly through binary assignment variables x_{o,j,t}, with constraints forcing full factory capacity, recipe eligibility, and exactly one factory per order. The reported experiments claim that the CB","pith_inferences":["Editorial inference: if a_{i,j,t} were made explicit as the sum of x over orders containing recipe i, the same MILP would become a textbook assignment formulation; the practical claim then partly rests on CBC's ability to exploit that structure—an extension the paper does not state.","Editorial inference: BAP's consecutive-day proxy is a temporal smoothing objective; the same formulation could be applied to other perishable, multi-site production settings (bakery, pharma, fresh produce) where day-to-day recipe or SKU mix stability drives waste.","Editorial inference: the reported inverse scaling effect suggests testing larger horizons or longer lead times may yield further error reduction, going beyond the 15-day frame the paper considers.","Editorial inference: using WMAPE global as the optimality benchmark is only valid when the factory-level decomposition can be permuted freely; as stated, Eq. (8)'s triangle inequality lower bound does not certify optimality unless site-level and global objectives coincide—an assumption the experiments appear to verify but the formulation does not prove."],"forward_implications":["If CBC solves BAP optimally at 100,000 orders in under two minutes, meal-kit operators can re-optimize allocations daily without buying commercial solvers.","Stable day-to-day recipe mixes make ingredient purchasing match actual production, lowering over-procurement and spoilage.","The proxy objective—minimizing consecutive-day changes—gives a computable route to the company's real goal of minimizing cumulative deviation from the final hard allocation at LD3.","The observed decrease of WMAPE with order volume means consolidation into larger allocation cycles should improve operational stability, not hurt it.","Under capacity shocks and order edits, an exact re-solve recovers optimality within days, whereas a greedy or ID-based carry-over rule accumulates error."],"supporting_citations":[{"why":"Internal company material that supplies the problem definition, box-structure figures, and the eligibility rules the model encodes.","marker":"[16]"},{"why":"CBC user guide that supplies the exact open-source solver used as the benchmark method.","marker":"[6]"},{"why":"Feasibility Pump heuristic used by CBC to obtain the initial integer-feasible solution.","marker":"[5]"},{"why":"Primal heuristics source for the Coefficient Diving method used to identify the optimal solution early.","marker":"[2]"},{"why":"Learning-to-dive variant of Coefficient Diving, cited as one of CBC's primal heuristic foundations.","marker":"[20]"},{"why":"Knapsack branch-and-bound background that underpins the exact B&B search and links BAP to bin-packing methods.","marker":"[18]"},{"why":"Vehicle Routing Problem reference from which the 2-Opt-inspired Iterative Targeted Pairwise Swap heuristic is adapted.","marker":"[25]"},{"why":"Tabu Search heuristic for vehicle routing that motivates the TS implementation.","marker":"[8]"},{"why":"Operations-management principle of aggregate forecasting used to interpret the decreasing error with larger order volumes.","marker":"[22]"},{"why":"Proxy-optimization framing that justifies minimizing consecutive-day changes as a surrogate for the harder cumulative objective.","marker":"[28]"}],"fun_headline_variants":["Exact solver for meal-kit box allocation at 100k orders","MILP optimizes meal kit factory recipes across 15 days","New BAP: temporal allocation solved in min for 100k orders","Meal kit delivery: MILP reduces recipe variation and waste"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that the recipe quantities a_{i,j,t} appearing in the objective are determined by the assignment variables x_{o,j,t}, even though the paper writes no equation linking them, and that each order consumes exactly one unit of factory capacity and contributes one unit of each of its recipes.","fun_headline_variants_meta":{"raw":{"variants":["Exact solver for meal-kit box allocation at 100k orders","MILP optimizes meal kit factory recipes across 15 days","New BAP: temporal allocation solved in min for 100k orders","Meal kit delivery: MILP reduces recipe variation and waste"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000196,"raw_usage":{"total_tokens":1173,"prompt_tokens":697,"completion_tokens":476,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":441,"completion_tokens_details":{"reasoning_tokens":412}},"tokens_in":441,"tokens_out":476,"duration_ms":5744,"temperature":1.0,"reasoning_tokens":412,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T00:00:43.094736+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the released code on a two-factory, two-order instance where both orders contain recipe 1 and are eligible for both factories, and print the objective for both feasible assignments. If the objective value is identical for both assignments, the model has no decision-variable dependence, and the reported optimality is vacuous. The GitHub repository named in the data-availability statement is the place to look.","supporting_citations":[{"cited_title":"(2024)Internal material","cited_arxiv_id":null,"evidence_quote":"Internal company material that supplies the problem definition, box-structure figures, and the eligibility rules the model encodes."},{"cited_title":"(2006)Primal heuristics for mixed integer programs","cited_arxiv_id":null,"evidence_quote":"Primal heuristics source for the Coefficient Diving method used to identify the optimal solution early."},{"cited_title":"and Toth, P","cited_arxiv_id":null,"evidence_quote":"Knapsack branch-and-bound background that underpins the exact B&B search and links BAP to bin-packing methods."},{"cited_title":"and Vigo, D","cited_arxiv_id":null,"evidence_quote":"Vehicle Routing Problem reference from which the 2-Opt-inspired Iterative Targeted Pairwise Swap heuristic is adapted."},{"cited_title":"1276-1290","cited_arxiv_id":null,"evidence_quote":"Tabu Search heuristic for vehicle routing that motivates the TS implementation."},{"cited_title":"and Sanders, N.R","cited_arxiv_id":null,"evidence_quote":"Operations-management principle of aggregate forecasting used to interpret the decreasing error with larger order volumes."},{"cited_title":"and Al-Kinani, A","cited_arxiv_id":null,"evidence_quote":"Proxy-optimization framing that justifies minimizing consecutive-day changes as a surrogate for the harder cumulative objective."}],"review_version":1}