{"id":"5ca125ff-7846-4f38-8681-4b90f7484cc8","arxiv_id":"2607.04497","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"Capacity-class cuts (RCI/multistar) lift CPTP B&C from 52 to 64 of 76 instances solved and shrink the tree >10×; comb, RGLM, fixing, propagation, and SPI add nothing measurable on the open HiGHS stack.","lead":"An open branch-and-cut solver for the capacitated profitable tour problem reimplements known cuts on HiGHS and shows, via ablation, that capacity-class cuts alone drive almost all gains. The released code and negative SPI result give a reusable baseline for routing pricing research.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified","rationale":"The paper promises a reusable open baseline plus a transparent component decomposition, not a new algorithm or a universal ranking of cut families. The strongest empirical claim is fully backed by the tables and the shipped code; the negative SPI result is cleanly documented. The instance-class limitation is acknowledged by the authors and correctly identified by the reader as the softest external assumption, but it is not load-bearing for the claim actually made. No further concern rises to the level that would move the verdict away from ACCEPT.","tokens_in":21541,"tokens_out":465,"duration_ms":4742,"concrete_test":"Re-run the full gating ladder of Table 3 from the committed ablation.csv / run_ablation.sh scripts on the released artifact (300 s limit, same 76 instances) and confirm that the solved-count jump 52\to64 and the SGM-node drop 2290\to214 still appear when capacity cuts are added, with comb/RGLM/SPI/fixing remaining flat. Any material reversal would falsify the headline claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is modest and tightly scoped: on this open HiGHS reimplementation of Jepsen et al., capacity-class cuts (RCI + multistar) drive essentially all of the measured gain over a GSEC-only baseline (52\to64/76 solved; SGM nodes 2290\to214), while comb, RGLM, reduced-cost fixing, bound-based propagation and SPI add nothing measurable on the 76 SPPRCLIB+Roberti instances (Table 3, §4.3). That claim is directly supported by the released ablation logs, cut-count tables, and runtime breakdown. The reader's weakest_assumption (representativeness of the undirected single-capacity pricing suite) is a real scope limit that the paper itself flags in §4.1, but it does not undercut the claim as stated—the authors do not assert transfer to arc-resource directed ESPPRC, and the open artifact is precisely the vehicle for others to test that. No internal inconsistency, separation bug, or unreproducible number is visible that would reverse the ablation ranking.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper re-implements the CPTP (and open s–t path) branch-and-cut of Jepsen et al. (2014) on a fully open HiGHS stack, adding bound-based preprocessing, domain propagation, and reduced-cost fixing, and releases the formulation, separators, and benchmark scripts. It claims no new method. The empirical contribution is a component study on 76 SPPRCLIB and Roberti pricing instances: capacity-class cuts (RCI and multistar) account for essentially the entire gain over a GSEC-only baseline (solved count 52→64/76; SGM nodes 2290→214), while comb, RGLM, reduced-cost fixing, bound-based propagation, and the SPI cut (a García-style node-precedence variant) add nothing measurable. Head-to-head with PathWyse labelling shows complementarity rather than dominance. All results are regenerated from released logs.","tokens_in":21746,"tokens_out":990,"duration_ms":9451,"significance":"If the ablation holds, the paper supplies a rare, fully open, rerunnable B&C baseline for CPTP/ESPPRC pricing and a clear ranking of which classical cut families actually pay off on modern open MIP software. The negative SPI result and the near-zero separation rates for comb/RGLM are useful negative knowledge. The open artifact (code, scripts, committed CSVs) is a concrete infrastructural contribution that others can extend to directed/arc-resource models. The claims are modest, tightly scoped, and directly supported by the released experiments; that is a strength, not a weakness, for a computational study.","major_comments":[{"comment":"§4.3 and Table 3 use a 300 s limit for the gating ablation, while the head-to-head comparison in §4.2 uses 3600 s. Several hard instances (e.g. M-n200-k17-12, E-n101-k14-158) sit near or at the 300 s wall in the ablation ladder. The ranking of capacity cuts vs. the rest is unlikely to reverse, but the paper should either re-run the ablation at 3600 s for the unsolved tail or explicitly justify that the 300 s ranking is stable (e.g. by reporting the same ranking on the subset solved under both limits).","section":null},{"comment":"§4.1 correctly notes that classical Beasley–Christofides and García arc-resource directed instances cannot be imported without a model change. The inertness claims for comb, RGLM, SPI, and reduced-cost fixing are therefore scoped to undirected single node-demand pricing instances. The abstract and conclusion should state this scope limit more prominently so that readers do not over-generalize the null results to the broader ESPPRC class the introduction targets.","section":null}],"minor_comments":[{"comment":"Table 7 reports mean wall-clock under the 3600 s limit for the default configuration; the ablation tables use 300 s. A short note that the component shares are qualitatively the same under both limits would avoid confusion.","section":null},{"comment":"§3.1: the SPI description (Held–Karp over |S|≤15) is clear, but a one-sentence remark that the separator never fires (Table 5) could be cross-referenced already in the method section so the negative result is foreshadowed.","section":null},{"comment":"Figures 2–3 are informative; adding the number of instances solved by each solver in the caption would make the performance profile self-contained.","section":null},{"comment":"The hyperplane branching machinery (§3.3) is described but never exercised. A single sentence that it is off by default and unused in the experiments would prevent readers from looking for results that are not there.","section":null},{"comment":"Minor typography: “opens–t path” in the conclusion should be “open s–t path”; a few long sentences in §1 and §4.3 would benefit from splitting for readability.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The paper is an honest, well-executed computational reimplementation study with an unusually complete open artifact. It is a good fit for a computational OR / mathematical programming journal that values reproducible baselines and negative results. I see no novelty-inflation or citation issues; the author correctly attributes the formulation and cuts to Jepsen et al. and García. Minor revision is appropriate; the two major points are scope/limit clarifications, not soundness problems."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is a clean, modest paper that does exactly what it says. No new algorithm. It reimplements the Jepsen et al. (2014) CPTP formulation and cut families on HiGHS, adds standard Aneja/García-style bound preprocessing and reduced-cost fixing, ships the whole thing open, and runs a gating ablation against PathWyse labelling on the 76 SPPRCLIB + Roberti pricing instances.\n\nWhat is actually new is the artifact and the decomposition. Capacity-class cuts (RCI + multistar) account for essentially the entire gain: GSEC alone solves 52/76; adding them reaches 64/76 and drops SGM nodes from 2290 to 214. Comb, RGLM, reduced-cost fixing, bound-based propagation, and the SPI variant of García’s node-precedence cuts add nothing measurable; SPI finds zero violated inequalities. Cut counts and runtime breakdowns make the mechanism transparent: without capacity cuts the separator floods with GSECs; with them the tree collapses and separation becomes cheap. LP + B&B dominate wall time (~98%). The complementarity with labelling on the hard tail is real and matches the older literature.\n\nSoft spots are real but scoped. The instance suite is undirected single node-demand pricing; classical arc-resource directed benchmarks do not fit the model, which the paper states plainly in §4.1. Ablation uses a 300 s limit while the head-to-head uses 3600 s. Default HiGHS branching is used throughout. None of that reverses the ranking on the suite they actually study. Math and separation look standard and carefully described; citations are appropriate; self-citations are to the software and an earlier technical report, not load-bearing circularity.\n\nThis is for people who build or use exact pricing / ESPPRC solvers and want a reusable open baseline plus a clear map of which components matter. It is not a theory paper and does not reorganize the field. I would send it to peer review without hesitation; the claims are modest, supported, and reproducible. Worth engaging if you care about open routing subproblem solvers or cut ablations.","headline":"Honest open reimplementation of Jepsen et al. with a clean ablation: capacity cuts do almost all the work; everything else is inert on this suite.","tokens_in":22381,"tokens_out":519,"would_cite":true,"duration_ms":5307,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C27","90C57","90C35"],"pacs":[],"model":"grok-4.5","headline":"Capacity cuts, not connectivity or fancy extras, drive almost all the gains in a branch-and-cut solver for capacitated profitable tours.","keywords":["profitable tour problem","elementary shortest path","branch-and-cut","resource-constrained shortest path","computational study","capacity cuts","open reproducible software"],"falsifier":"Re-run the same gating ablation on a larger or more strongly capacity-binding set (or on a directed arc-resource reformulation of classical resource-constrained shortest-path benchmarks) and check whether comb, rounded-GLM, SPI, or reduced-cost fixing begin to separate violated cuts and shrink the tree; if they do, the inertness claim fails.","tokens_in":22407,"feed_emoji":"✂️","tokens_out":1010,"duration_ms":9591,"temperature":0.7,"pith_summary":"This paper releases an open, fully reproducible branch-and-cut solver for the capacitated profitable tour problem and its open path version, rebuilt on a free mixed-integer programming stack. The author claims no new algorithm: the formulation and cut families follow earlier work. The real contribution is a careful component study that shows which pieces actually matter. Adding capacity-class cuts (rounded capacity and multistar inequalities) to a connectivity-only baseline is what moves the needle: solved instances rise from 52 to 64 of 76, and the search tree shrinks more than tenfold. Comb inequalities, rounded generalized multistar cuts, reduced-cost fixing, bound-based propagation, and a shortest-path-incompatibility cut add essentially nothing. The same study shows the branch-and-cut is complementary to dynamic-programming labelling: each method closes hard instances the other cannot. The open artifact and the negative results are meant as a reusable baseline so others need not rediscover what fails.","feed_headline":"Capacity cuts alone shrink the search tree tenfold","feed_subtitle":"Open branch-and-cut study shows comb, fixing and SPI add nothing on 76 instances","key_machinery":"Capacity-class cuts (rounded capacity inequalities and multistar/generalized large multistar inequalities) separated from a shared Gomory–Hu cut tree on the fractional support graph. These node-coupled inequalities tighten the knapsack-plus-circuit polytope; the ablation isolates them as the only family that measurably reduces the search tree and increases the number of solved instances.","core_discovery":"On a common modern open stack, capacity-class cuts account for essentially the entire benefit of the branch-and-cut: they lift the number of instances solved from 52 to 64 of 76 and shrink the search tree more than tenfold, while comb cuts, rounded generalized large multistar cuts, reduced-cost fixing, bound-based propagation, and the shortest-path-incompatibility cut contribute nothing measurable. The branch-and-cut and dynamic-programming labelling remain complementary rather than one dominating the other.","pith_inferences":["Because the classical directed arc-resource benchmarks cannot be imported without a model change, the paper’s inertness conclusions are currently untested outside the undirected node-demand regime; an arc-resource reimplementation would be the natural next falsification test.","The near-zero separation counts for comb, RGLM and SPI suggest the failure is one of separator design relative to the incumbent, not necessarily of theoretical strength; alternative separation heuristics might revive them.","If capacity-class cuts remain decisive on larger instances, pricing engines inside column-generation vehicle-routing solvers could profitably expose the same cut families as user cuts rather than relying solely on labelling."],"forward_implications":["A practical default for this problem class can keep only generalized subtour elimination plus rounded-capacity and multistar cuts; the remaining families can stay optional and nearly free.","Open, scriptable baselines make future component studies and extensions (per-arc resources, directed graphs) directly comparable instead of locked inside proprietary stacks.","Negative results on SPI and related node-precedence-style cuts discourage re-implementation of those families for the same instance regime.","Complementarity with labelling implies that hybrid or portfolio solvers can close more of the hard tail than either method alone."],"fun_headline_variants":["Capacity cuts alone lift CPTP solves from 52 to 64 of 76","Only capacity cuts drive tenfold search-tree shrink on CPTP","Open B&C: capacity cuts account for every measurable gain","Comb, fixing and SPI cuts add zero on 76 CPTP instances","Capacity-class cuts shrink tree >10x; rest add nothing"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The 76 undirected, single-capacity pricing instances used for the ablation are representative enough that the conclusions about which cuts and fixing routines help or fail will hold for the wider class of capacitated profitable-tour and elementary shortest-path problems.","fun_headline_variants_meta":{"raw":{"variants":["Capacity cuts alone lift CPTP solves from 52 to 64 of 76","Only capacity cuts drive tenfold search-tree shrink on CPTP","Open B&C: capacity cuts account for every measurable gain","Comb, fixing and SPI cuts add zero on 76 CPTP instances","Capacity-class cuts shrink tree >10x; rest add nothing"]},"model":"grok-4.5","effort":"low","cost_usd":0.004662,"raw_usage":{"total_tokens":1449,"prompt_tokens":912,"num_sources_used":0,"completion_tokens":96,"cost_in_usd_ticks":46620000,"prompt_tokens_details":{"text_tokens":912,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":441,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":912,"tokens_out":96,"duration_ms":4149,"temperature":1.0,"reasoning_tokens":441,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-11T18:34:38.671944+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Re-run the same gating ablation on a larger or more strongly capacity-binding set (or on a directed arc-resource reformulation of classical resource-constrained shortest-path benchmarks) and check whether comb, rounded-GLM, SPI, or reduced-cost fixing begin to separate violated cuts and shrink the tree; if they do, the inertness claim fails.","supporting_citations":[],"review_version":1}