{"id":"27d3bed3-a079-43b1-b1fc-44a0ba520b90","arxiv_id":"1908.10293","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":2.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A linear volume discount added to a classic lot-sizing model changes the optimal plan and cuts total cost in a synthetic six-period example, but no baseline without the discount is reported.","lead":"This undergraduate thesis models a lot-sizing problem where unit production cost falls linearly with lot size, mimicking learning and forgetting, and solves a six-period example in GAMS. It reports large cost savings from the assumed discount, but the result follows directly from the assumed cost structure rather than from a validated learning effect.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline effect is an artifact of the assumed per-lot discount: Eq. (8) has no cumulative learning/forgetting state, so the 30% reduction does not demonstrate real learning-curve effects.","rationale":"The reader correctly identifies the linear learning curve and local-optimality caveat as weak points. My concern is more direct: the implemented Eq. (8) is not a learning curve in the usual cumulative-output sense; it is a per-period quantity discount with a full reset, so the paper's central claim is closer to a tautology than to an empirical or independently supported finding. The Section 5 percentage also appears to misstate the discount by a factor of 10, which directly affects the headline quantitative result. This strengthens the conditional verdict rather than overturning it: the model is coherent as a small quantity-discount lot-sizing exercise, but it cannot support the claim that learning/forgetting effects are important without a baseline and a genuine state-dependent learning curve. The proposed test (zero-discount baseline plus a cumulative-learning variant) would settle whether any nontrivial learning/forgetting dynamics are needed. The appendix code mismatch also hurts reproducibility, but it is secondary. Agreement with the reader is partial because the reader focused on linearity rather than on the absence of cumulative learning/forgetting state.","tokens_in":17799,"tokens_out":7351,"duration_ms":73294,"concrete_test":"Re-run all 18 parameter combinations with x=0 and, separately, replace Eq. (8) with a cumulative learning curve that includes a forgetting reset, e.g., uc_i = 100 - x*(cumulative production before period i) with a parameter for the fraction of learning lost after idle periods. If the x=0 baseline already explains the production plans and the 30% cost gap through the mechanical discount, or if the cumulative model changes the optimal plans and cost reductions significantly, then the reported learning/forgetting effect is an artifact of the per-lot discount assumption.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim in Section 5 (\"Um desconto de 0,1% por produto do custo inicial de produção gerou reduções de custo de 30%\") is not supported by the model as specified. In Eq. (8), uc_i = 100 - x*p_i, the unit cost in period i depends only on that period's production quantity, with the intercept reset to 100 every period. There is no cumulative production term and no state variable for idle time; hence the \"learning\" is a within-period quantity discount and \"forgetting\" is an inflexible full reset. The reported 30% reduction is therefore a direct mathematical consequence of choosing x=0.01 and capacity 5000: the maximum discount is 0.01*5000 = $50 per unit (50% of the $100 base cost), and production cost dominates total cost. Moreover, the final text says 0.1%, but the table uses x=0.01, which is 0.01% of $100; the claimed percentage is off by a factor of 10. Because no independent learning-curve data or a true x=0 baseline is reported, the \"importance of learning and forgetting\" is not demonstrated; the paper demonstrates only that adding a large assumed price discount to a cost-dominated objective lowers the optimum. The acknowledged linearity (Section 2.4) and local-optimality caveat (Section 4) further weaken the quantitative claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript, an undergraduate monograph in Portuguese, proposes an extension of a single-machine, single-product lot-sizing model in which the unit production cost is reduced by a linear term proportional to the lot size in each period (uc_i = 100 - x * p_i). The authors solve the resulting mixed-integer quadratic program with GAMS over a six-period deterministic demand scenario, varying inventory holding costs, setup costs, and the discount coefficient x. They report that including the discount changes the optimal lot sizes and reduces total costs by up to about 30%, and they conclude that learning and forgetting effects are important and should be considered in lot sizing. The paper also notes that the quadratic formulation may yield only local optima and that the assumed linear learning curve is a simplification.","tokens_in":18058,"tokens_out":3220,"duration_ms":33038,"significance":"If the central claim were supported, the paper would provide a simple quantitative illustration that lot-sizing decisions change when production cost declines with cumulative output, and it would complement the existing literature on learning and forgetting in lot sizing. The GAMS model is transparent and easily reproducible, and the parameter sweep across holding costs, setup costs, and discount levels is a reasonable exploratory design. However, the model as specified does not implement learning or forgetting: the unit cost depends only on the current period's production quantity, not on cumulative experience or idle time, so the reported cost savings are a within-period quantity discount. The paper's headline quantitative claim also contains a numerical inconsistency. The core contribution, demonstrating the importance of learning and forgetting, is therefore not established.","major_comments":[{"comment":"The model's learning/forgetting mechanism is given by uci = 100 - x*pi, which depends only on the production quantity in period i and resets to a base cost of 100 in every period. There is no state variable for cumulative production, no dependence on previous periods' output or idle time, and no decay term that would represent forgetting. Consequently, the model is a within-period quantity-discount model, not a learning-curve model as defined in Section 2.4. The conclusion in Section 5 that the results demonstrate 'que os efeitos de learning e forgetting são sim importantes' is not supported by the equations actually solved.","section":"Section 3.3.2, Eq. (8)"},{"comment":"The statement that 'Um desconto de 0,1% por produto do custo inicial de produção gerou reduções de custo de 30%' is numerically inaccurate: the largest tested value x=0.01 corresponds to 0.01% of the base cost of $100, not 0.1%. Moreover, the 30% cost reduction is computed relative to the x=0.0001 (nearly zero discount) scenario, not relative to a true baseline with x=0, and it is a direct consequence of the assumed linear discount reaching up to 50% of unit cost at the capacity bound of 5000 units. A baseline run without the discount is needed to isolate the effect of the learning/forgetting term; without it, the quantitative importance claim is overstated.","section":"Section 5 and Table 3"},{"comment":"The paper acknowledges in Section 4 that the quadratic results may represent only local optima, and the GAMS code in Appendix A fixes p('5')=1400 and solves with optcr=0, which does not guarantee global optimality for the nonconvex MIQCP. Since the reported savings are differences between possibly local solutions of different scenarios, the magnitude of the claimed effect is not reliable. A comparison with a known optimal solution for the linear (x=0) problem or a global-optimization certificate would be necessary to support the reported cost reductions.","section":"Section 4 and Appendix A"}],"minor_comments":[{"comment":"The abstract and Section 2.4 describe a 'learning curve', but the accepted model is explicitly linear; this discrepancy should be stated more clearly as a deliberate simplification with known limitations, not just in passing.","section":"Abstract and Section 2.4"},{"comment":"The list of symbols defines xi as the discount, but the model text and Table 2 use x without a subscript; the notation should be made consistent.","section":"List of Symbols and Table 2"},{"comment":"The sentence before Eq. (8) says 'A equação número 7 mostra onde o efeito de learning e forgetting foi implementado', but Eq. (7) is the capacity constraint; the new discount equation is Eq. (8). The equation reference should be corrected.","section":"Section 3.3.2"},{"comment":"The final paragraph correctly mentions the possibility of a more realistic logarithmic learning curve in future work, but this caveat should also appear where the linear assumption is first introduced, not only in the conclusions.","section":"Section 5"},{"comment":"Several references are incomplete (e.g., missing page ranges or journal identifiers), and one reference to Pinedo (2010) is cited in the text as 'Pinedo (2010)' while the bibliography lists 'Pinedo M. Lot-sizing... 2002'; the citation style should be harmonized.","section":"References"}],"recommendation":"reject","confidential_remarks":"The manuscript is an undergraduate monograph with a transparent and reproducible GAMS implementation, but the central modeling choice does not implement learning or forgetting as claimed, and the quantitative 'importance' result is an artifact of the within-period discount and the lack of a baseline. In my view, the issue is not merely presentational: correcting it would require replacing Eq. (8) with a cumulative-learning state and a forgetting mechanism, which is beyond a revision of the current text. The paper also has a factor-of-ten error in the headline percentage. I recommend rejection rather than major revision, though a rewritten paper with a proper learning/forgetting model and a true x=0 baseline could be a legitimate future submission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a straightforward undergraduate monograph that adds a linear volume discount to a standard lot-sizing MILP and solves a small GAMS instance. The math is fine, but the central claim that it demonstrates learning and forgetting effects is not supported.\n\nWhat is genuinely there: a clean statement of a single-product, single-machine lot-sizing model with capacity, setup, and holding costs, plus the unit-cost equation uc_i = 100 - x*p_i. The paper is honest about the linearity assumption (Section 2.4) and about local optimality (Section 4). It cites the right survey literature (Cheng et al. 2004, Biskup 2008), and as a classroom exercise in building a quadratic mixed-integer model in GAMS it is coherent.\n\nThe soft spots are serious, though. Equation (8) has no cumulative production or idle-time state. Unit cost in period i depends only on that period's p_i, so the \"learning\" is a within-period quantity discount and \"forgetting\" is an automatic full reset at every period boundary. That is not a learning curve in the sense used by the surveys the paper cites; it is a volume discount. The quantitative conclusion in Section 5 — that a 0.1% discount caused 30% cost reductions — is an arithmetic consequence of the assumed x and capacity. With x=0.01 and p=3600, the unit cost drops to 64, and production cost dominates the objective. Running the same model with x=0 would give a baseline; the paper does not report one, so the claimed \"importance\" is never measured against the no-learning case. There is also a factor-of-ten discrepancy between the text (0.1%) and the x=0.01 coefficient (0.01% of base cost on the first unit), and the appendix code does not match the main results: setup costs are 4000 in the code while the table uses 2000/1500 patterns, p.fx('5')=1400 is hard-coded, and the GAMS output shows a different production plan and a total cost of 10,600. So reproducibility is poor.\n\nNet: the model is coherent but the demonstration does not support the paper's headline. This is an honest student project, not a deceptive one; it simply overinterprets a parameter sweep. I would not put it in front of a referee as a research contribution. If the authors ever revisit it, adding an x=0 baseline, a genuinely cumulative learning variable, and matching code would be the way to go.","headline":"An honest undergraduate monograph whose \"learning and forgetting\" is modeled as a within-period linear quantity discount, making the headline 30% savings an artifact of the assumed coefficient rather than a demonstrated learning-curve effect.","tokens_in":18611,"tokens_out":4271,"would_cite":false,"duration_ms":38561,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90B30","90C20","90C11"],"pacs":[],"model":"deepseek-v4-flash","headline":"A 0.1% learning discount cut modeled lot-sizing costs by 30%","keywords":["lot-sizing","learning curve","forgetting effect","quadratic programming","mixed-integer quadratically constrained programming","economies of scale","variable demand","production planning"],"falsifier":"Run the same six-period problem with the discount applied to cumulative production instead of the current period's batch, and solve to proven global optimality; if the 30% saving shrinks or disappears, the linear per-batch discount is doing the work.","tokens_in":17548,"feed_emoji":"📉","tokens_out":6449,"duration_ms":60237,"temperature":0.7,"pith_summary":"This paper tries to establish that the learning and forgetting effects, usually ignored in lot-sizing models, materially change both the optimal production plan and the total cost. It extends a standard linear lot-sizing formulation by making the unit production cost fall linearly with the size of the period's lot, then solves the resulting mixed-integer quadratic program for a six-period, single-product case. Across 18 scenarios, even a per-unit discount of one-tenth of a cent (0.01% of unit cost) shifted production quantities and saved about $10,000, while a 0.1% discount produced total-cost reductions of about 30%. The paper's point is that economies of scale from learning deserve to be included, not treated as negligible.","feed_headline":"A 0.1% learning discount cut modeled lot-sizing costs by 30%","feed_subtitle":"Even minimal learning and forgetting effects change optimal production plans and total cost, a six-period model shows.","key_machinery":"The load-bearing mechanism is a linear learning/forgetting discount embedded in the unit production cost: $uc_i = 100 - x p_i$, where $x$ is the per-unit discount and $p_i$ is the lot size in period $i$. This replaces the fixed per-unit cost of the classical model with one that rewards larger lots, and the product $uc_i \\cdot p_i$ becomes $100 p_i - x p_i^2$, which is why the model is a mixed-integer quadratic program. The forgetting side is represented by resetting the discount each period, so learning does not carry across periods. The machinery's job is to turn the learning curve into a quantity discount that the optimizer can trade against setup and holding costs.","core_discovery":"The central discovery is that adding a per-period linear learning/forgetting discount to the variable production cost changes the lot-sizing solution enough to matter. In the model, unit cost in period $i$ becomes $uc_i = 100 - x p_i$, so the production term in the objective becomes $(100 p_i - x p_i^2)$, making the problem quadratic. With the GAMS solver handling the mixed-integer quadratically constrained program, the author found that the discount coefficient $x = 0.01$ (0.1% of the base unit cost) produced total-cost reductions around 30% compared with the no-discount baseline, and even $x = 0.001$ produced roughly $10,000 in average savings. The paper also reports that variable setup costs helped mainly through strategic timing of production, saving about $1,000, while holding-cost differences could be offset by a slightly larger discount. The author states these solutions are not guaranteed globally optimal because quadratic models may return a local minimum.","pith_inferences":["I read the model's 'forgetting' as a period reset: since unit cost depends only on the current period's $p_i$, the formulation behaves like a quantity discount rather than a cumulative learning curve; switching to cumulative output would likely change which periods produce.","The reported 30% figure is tied to this six-period demand pattern; testing on longer horizons, different seasonal peaks, or capacity-constrained settings would show how general the magnitude is.","A natural extension is to compare the linear discount against a logarithmic learning curve with explicit forgetting decay; if results are similar, the conclusion is robust, and if not, curvature is decisive."],"forward_implications":["Optimal lot sizes shift toward fewer, larger production runs when the discount coefficient is high enough.","Total cost falls by amounts that are large relative to the discount: a 0.1% per-unit discount yields about a 30% total-cost reduction in the studied case.","Ignoring learning and forgetting overstates cost and can misidentify when to produce.","Variable setup costs produce smaller savings, around $1,000, but still improve the plan.","The linear-discount formulation requires quadratic programming, so linear lot-sizing solvers are insufficient."],"supporting_citations":[{"why":"Supplies the base linear lot-sizing model that the paper extends with the learning/forgetting discount.","marker":"Arenales et al (2007)"},{"why":"Survey of lot-sizing with time-dependent processing times, used to justify considering learning effects.","marker":"Cheng et al (2004)"},{"why":"Recent review cited as evidence that learning and forgetting are being incorporated into lot-sizing and scheduling research.","marker":"Biskup (2008)"},{"why":"Provides the learning-curve definition used to justify lower unit costs with higher output.","marker":"Hill (2011)"},{"why":"Prior single-machine lot-sizing model with learning effects that this work extends.","marker":"Li-Yan et al (2009)"},{"why":"Two-machine flow-shop lot-sizing with linear forgetting, supporting the linearity assumption used here.","marker":"Shiau et al (2007)"}],"fun_headline_variants":["0.1% learning discount slashes lot-sizing costs 30%","Minimal learning effects cut lot-sizing cost by 30%","Even tiny learning discounts reshape lot-sizing plans","Learning and forgetting effects matter in lot-sizing: 30% savings","Quadratic lot-sizing: 0.1% discount yields 30% cost cut"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole result depends on the assumption that learning and forgetting lower the unit cost in a straight line based only on the size of the current batch; real learning curves are usually curved and build on cumulative experience.","fun_headline_variants_meta":{"raw":{"variants":["0.1% learning discount slashes lot-sizing costs 30%","Minimal learning effects cut lot-sizing cost by 30%","Even tiny learning discounts reshape lot-sizing plans","Learning and forgetting effects matter in lot-sizing: 30% savings","Quadratic lot-sizing: 0.1% discount yields 30% cost cut"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000425,"raw_usage":{"total_tokens":2154,"prompt_tokens":894,"completion_tokens":1260,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":510,"completion_tokens_details":{"reasoning_tokens":1170}},"tokens_in":510,"tokens_out":1260,"duration_ms":9640,"temperature":1.0,"reasoning_tokens":1170,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:19:39.232140+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same six-period problem with the discount applied to cumulative production instead of the current period's batch, and solve to proven global optimality; if the 30% saving shrinks or disappears, the linear per-batch discount is doing the work.","supporting_citations":[{"cited_title":"e YANASSE H","cited_arxiv_id":null,"evidence_quote":"Supplies the base linear lot-sizing model that the paper extends with the learning/forgetting discount."},{"cited_title":"and LIN B.M.T","cited_arxiv_id":null,"evidence_quote":"Survey of lot-sizing with time-dependent processing times, used to justify considering learning effects."},{"cited_title":"A state -of-the-art review on lot -sizing, European Journal of Operational Research, vol 188, no","cited_arxiv_id":null,"evidence_quote":"Recent review cited as evidence that learning and forgetting are being incorporated into lot-sizing and scheduling research."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the learning-curve definition used to justify lower unit costs with higher output."},{"cited_title":"and EN -MIN F","cited_arxiv_id":null,"evidence_quote":"Prior single-machine lot-sizing model with learning effects that this work extends."},{"cited_title":"R., LEE W","cited_arxiv_id":null,"evidence_quote":"Two-machine flow-shop lot-sizing with linear forgetting, supporting the linearity assumption used here."}],"review_version":1}