{"id":"c7283716-eb74-4b93-bd38-3058c1bb4ccb","arxiv_id":"2507.13005","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"On stylized transport networks, manual link-ordering strategies outperform greedy and random strategies in the directness-coverage trade-off, with centrality-based greedy best for directness and coverage-based greedy best for coverage but poor for directness.","lead":"This paper studies the order in which links of a planned transport network should be built, comparing random, greedy, and human-designed strategies on stylized city grids. It finds that human-designed sequences beat greedy ones, which beat random ones, but with different strategies favoring directness or coverage.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Manual orders C and D are asserted, not proven, to be globally optimal; the 'manual beats greedy' ranking rests on this unverified premise.","rationale":"The stress-test pass confirms the reader's main reservation. The paper is careful and reproducible: strategies are precisely defined, figures report per-run variation, open-source code is provided, and the headline greedy-versus-random comparison is well supported. The vulnerable point is not the simulations but the interpretive frame: Sections 2.3 and 3.2 treat manual orders C and D as globally optimal upper bounds for coverage and directness AUC. No proof, certificate, or exhaustive check is supplied, and the authors explicitly disclaim optimality for the other three networks. Because these orders are the reference points that put 'manual' on the Pareto front and motivate the abstract's ranking, the claim is load-bearing. The need is not for a new experiment but for a justification of optimality, or a softened claim that these are strong manual heuristics rather than proven upper bounds. I also note a smaller internal overstatement: the abstract says coverage-based greedy strategies reach maximum global coverage 'as fast as possible', whereas Section 3.3 reports manual C needs 48 links and the Coverage greedy needs 58, so that phrase is at best true only among greedy strategies. Neither issue changes the reader's conditional verdict, but both should be addressed before the findings are used as design guidance.","tokens_in":20366,"tokens_out":15439,"duration_ms":195934,"concrete_test":"Run an exact dynamic program over connected edge subsets for the 4x4 and, if feasible, 4x5 grids (24 and 31 links), computing the true maximum-AUC coverage and directness orders under the same seed-link and connectedness rules, then compare these optima to the manual C/D construction applied to the same small grids. A counterexample would falsify the asserted global optimality for the regular grid. As a direct 10x10 check, run simulated annealing initialized at manual C/D using the authors' released orderbike code; any improvement over C/D would refute the claimed upper bound, while repeated failure to improve would make the premise more credible but would still not constitute a proof.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.3 states 'We are able to design globally optimal manual orders for the regular grid', and Section 3.2 uses this to place manual orders C and D on the Pareto front and to conclude that manual strategies outperform greedy ones. No proof, certificate, or exhaustive check is provided, and the authors explicitly concede that global optimality cannot be argued for the other three test networks. This premise is load-bearing for the headline 'manual beats greedy': if C and D are merely well-tuned heuristics, the result reduces to 'a hand-designed heuristic can beat greedy', not 'manual strategies attain the global optimum', and a better manual or hybrid greedy order could shift the ranking. The optimality is nontrivial: for coverage AUC the objective is a weighted sum of submodular increments, so greedy-by-largest-immediate-increment is not guaranteed to be globally optimal; for directness AUC the metric is highly nonlinear and depends on which nodes have entered the growing component. The manual-vs-greedy margin is also partly due to a composite rule (manual C switches to manual H after reaching full coverage), whose optimality is not separately established. A smaller internal overstatement is that the abstract says coverage-based greedy strategies reach maximum global coverage 'as fast as possible', while Section 3.3 reports manual C reaches full coverage in 48 links versus 58 for the Coverage greedy.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the order in which links of a fixed connected planar transport network should be built so that the growing network is functional early. It defines coverage and directness metrics, compares random ordering, six greedy strategies (coverage, adaptive coverage, directness, relative directness, closeness, betweenness), and three hand-designed manual orders (H, C, D) on a 10x10 grid and three other stylized networks. Performance is summarized by normalized area under the curve (AUC). The main findings are that random orders are far from the Pareto front; greedy strategies vary widely, with centrality-based strategies doing well on directness but poorly on coverage and coverage-based strategies doing the opposite; manual orders C and D sit on the Pareto front and outperform greedy strategies on the grid; and additive strategies generally outperform subtractive ones. The authors report robustness of the main ordering to grid distortions and to exponential discounting of early growth.","tokens_in":20624,"tokens_out":5926,"duration_ms":65341,"significance":"If substantiated, the paper provides a useful quantitative ranking of link-ordering strategies for transport network growth, with direct relevance to bicycle network planning. Its strengths include open reproducible code, a systematic comparison across four stylized urban topologies, and explicit robustness checks (distorted grids, exponential discounting). The central caveat is that the headline 'manual beats greedy' claim rests on the asserted but unproven global optimality of two hand-designed orders on the regular grid. This makes the paper's conclusions conditional rather than fully established, but the underlying empirical comparison of random and greedy strategies is sound and reproducible.","major_comments":[{"comment":"The claim that manual orders C and D are globally optimal is asserted without proof. Section 2.3 states 'We are able to design globally optimal manual orders for the regular grid,' and Section 3.2 treats the Pareto-front placement of C and D as 'expected since we designed them to be optimal strategies.' However, no certificate, exhaustive check, or formal argument is provided, and the same section concedes that global optimality cannot be argued for the other test networks. This premise is load-bearing: the conclusion that manual strategies beat greedy ones depends on C and D being true optima rather than merely well-tuned heuristics. The composite rule in manual order C (switching to order H after full coverage) is also not separately justified as optimal. Please either supply a proof or verification (for example, a branch-and-bound or exhaustive check on a smaller grid, or a comparison against a stronger metaheuristic such as simulated annealing) or rephrase the claims as 'hand-designed heuristic strategies' and soften the 'upper bound' language throughout.","section":"2.3, 3.2"},{"comment":"The abstract says that coverage-based greedy strategies 'can achieve maximum global coverage as fast as possible,' but Section 3.3 reports that manual order C reaches full coverage with 48 links (27% of 180) whereas the additive Coverage strategy needs 58 links (32%). This is internally inconsistent unless 'as fast as possible' is explicitly restricted to greedy strategies. The same wording appears in the Discussion. Please clarify or correct the claim.","section":"Abstract, 3.3, 4"},{"comment":"The abstract's claim that 'manual strategies fare better than greedy strategies on both metrics' is stronger than the reported data. Section 3.2 itself notes that the Directness greedy strategy has slightly more AUC of directness than manual order H and that the Closeness strategy has slightly more AUC of coverage than manual order D. Moreover, manual orders are constructed only for the regular grid; Section 3.7 runs only random and greedy strategies on the other three networks. The cross-network claim should be restricted to the greedy-versus-random comparison, and the manual-versus-greedy claim should be limited to the grid or to the specific manual strategies C and D on their respective target metrics.","section":"Abstract, 3.2, 3.7"},{"comment":"The statement that the average random order is 'at least six standard deviations away from the front for each axis' needs a precise definition of distance to the Pareto front. The front is a curve in the (AUC of directness, AUC of coverage) plane, so 'each axis' is ambiguous. The subsequent claim that about 500 million random trials would be needed also assumes normality for the distance statistic, while the Shapiro-Wilk test is reported only for the coverage marginal. Please replace this with a directly defined statistic, for example the fraction of random runs falling within a fixed neighborhood of the Pareto front, and report its distribution.","section":"3.1"}],"minor_comments":[{"comment":"The expression for the buffered link area appears to contain a typographical or formatting error ('2ld + 21 2πd2'); the intended formula is presumably 2ld + πd^2. Please check and correct.","section":"2.2, Eq. (1)"},{"comment":"The adaptive-coverage threshold (0.1), the initial buffer size (300 m), and the halving rule are chosen heuristically. Since Section 3.5 shows that the choice of buffer size materially affects the results, a brief sensitivity analysis for these parameters, or at least an explicit statement that they are calibrated choices, would strengthen the paper.","section":"2.2, 3.5"},{"comment":"The text says that the additive Coverage strategy reaches full coverage with 58 links (32%) and then states that 'every link added after the first 59 is chosen at random.' The off-by-one discrepancy should be reconciled.","section":"3.3"},{"comment":"The caption of Fig. 6 says the Pareto front is 'computed for all considered orders,' but it is not clear whether this includes all single runs or only per-strategy average curves. Please define this precisely.","section":"3.2, Fig. 6"},{"comment":"The software repositories are cited by URL without a version identifier or DOI; adding a versioned release or archive would improve reproducibility.","section":"Declarations"}],"recommendation":"major_revision","confidential_remarks":"The empirical core of the paper—the reproducible simulations comparing random and greedy strategies—is solid and publishable. The main obstacle is the unproven optimality of the manual orders C and D, which the authors themselves acknowledge is unavailable for the non-grid networks. If the authors can supply a verification for the grid or convincingly rescope the claims to 'hand-designed heuristics,' the paper would be acceptable. The paper fits the journal's scope, and the open-code practice is a genuine strength."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a genuinely useful systematic comparison of link-ordering strategies for growing connected planar networks, with open code and clear figures. The headline empirical ranking—manual > greedy > random in directness-coverage AUC—is plausible but rests on a weaker foundation than the paper admits.\n\nWhat's new: previous work like Szell et al. 2022 looked at bicycle network growth, but a unified AUC framework comparing additive and subtractive greedy objectives (coverage, adaptive coverage, directness, relative directness, closeness, betweenness) plus human-designed orders across four stylized topologies is new. The finding that random orders are hopelessly far from the Pareto front, and the explanation of why directness greedy falls into local optimum traps, are solid and well illustrated. The code is available and the reproducibility bar is met.\n\nThe soft spot is real. Section 2.3 asserts that manual orders C and D are globally optimal for coverage and directness on the regular grid, but no proof or certificate is given. The claim is load-bearing: 'manual beats greedy' uses C and D as upper bounds. Since coverage is submodular, greedy-by-largest-increment is not guaranteed optimal, and directness is highly nonlinear; it is not obvious that these hand-designed orders are truly optimal. The paper even concedes it cannot argue global optimality for the other networks. If C and D are just good heuristics, the margin from greedy shrinks and the headline becomes 'a well-tuned heuristic can outperform greedy,' which is weaker.\n\nA smaller issue: the abstract says coverage-based greedy strategies reach maximum global coverage as fast as possible, but Section 3.3 reports manual C does it in 48 links versus 58 for Coverage greedy. That is an overstatement. The buffer size is a modeling choice, and the authors acknowledge sensitivity; minor.\n\nThe paper is honest about its stylized scope, and the central empirical content—random is far from Pareto, directness greedy gets trapped, additive beats subtractive—does not depend on the unproven optimality. That content is reproducible and likely robust.\n\nThis paper deserves a serious referee. The main fix is to either prove the optimality of C and D, or reframe the claim as 'manual heuristics can beat greedy' and soften the upper-bound language. I'd bring it to a reading group and would cite the framework, but I would not hang a strong 'manual beats greedy' conclusion on it as currently stated.","headline":"Useful systematic comparison of network growth strategies, with open code, but the 'manual beats greedy' claim rests on unproven optimality of two hand-designed orders.","tokens_in":21142,"tokens_out":1954,"would_cite":true,"duration_ms":21078,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Hand-built link orders beat greedy and random growth on both metrics.","keywords":["link ordering","transport network growth","directness","coverage","greedy algorithm","planar networks","bicycle networks","network evolution"],"falsifier":"For the 10×10 grid, solve the link-ordering problem exactly with an integer program or branch-and-bound search that maximizes the area under the curve of coverage or directness over all permutations of the 180 links, and compare those exact optima with the AUC values reported for manual orders C and D; if either exact optimum differs from the reported manual value, the claim that manual orders are upper bounds and therefore beat greedy strategies would be falsified.","tokens_in":20176,"feed_emoji":"🚲","tokens_out":7548,"duration_ms":80704,"temperature":0.7,"pith_summary":"The paper asks: given a planned final transport network, in what order should links be built so the network is already useful while under construction? It compares random, greedy, and human-designed link-ordering strategies on stylized connected planar networks, scoring each growth curve by area under the curve for directness and coverage. The central finding is a ranking: hand-designed orders achieve the best trade-off, greedy optimization is intermediate but metric-dependent, and random orders are so far from the efficient frontier that brute-force random search would need hundreds of millions of trials. Greedy strategies aimed directly at directness fall into local optimum traps, while coverage-maximizing greedy reaches full coverage quickly but then drifts to random directness. The result matters because bicycle and other sustainable transport networks are built incrementally over decades, so the order of construction determines how early the network becomes usable.","feed_headline":"Hand-built link orders beat greedy and random growth on both metrics","feed_subtitle":"The order in which links are built decides whether a transport network is useful before it is finished.","key_machinery":"The central object is a link ordering strategy: an algorithm that takes the final planar network and outputs a permutation of its links, and the paper evaluates each strategy by the area under the curve (AUC) of its coverage and directness trajectories over normalized built length. Directness is the average ratio of Euclidean to shortest-path distance over all node pairs; coverage is the area of buffered built links. The comparison engine is a set of additive greedy algorithms—at each step adding the link that maximizes the chosen metric while keeping the network connected, with random tie-breaking—plus two hand-designed orders that the authors argue are globally optimal for coverage and directness on the regular grid, and a hierarchical heuristic. AUC in coverage-directness space is what converts growth curves into a single ranking and lets the paper identify the Pareto front.","core_discovery":"On the paper's own terms, the central claim is that for connected planar networks with a fixed final link set, the choice of growth strategy creates a nontrivial directness-coverage trade-off: manual orders C and D, hand-designed as upper bounds for coverage and directness on a regular grid, dominate greedy strategies on both AUC metrics; greedy directness, closeness, and betweenness strategies sit on or near the Pareto front for directness but sacrifice coverage; and random ordering is effectively a lower bound, six standard deviations from the Pareto front. Coverage-based greedy strategies attain maximal coverage AUC as fast as possible but perform no better than random on directness, and directness-based greedy gets trapped in local optima because long cycles are hard to complete. The same qualitative ranking holds across a grid with a diagonal, a three-bridges network, and a radio-concentric network, and the ranking persists under grid distortion and exponential discounting of early growth stages.","pith_inferences":["The paper only tests stylized topologies with homogeneous node distributions; a natural extension is to empirical street networks with nonuniform population and destinations, where coverage should be weighted by actual demand rather than raw buffer area, and the paper notes the all-to-all directness computation would scale as O(L^5).","The 'manual beats greedy' result suggests a step-dependent objective—start with coverage-greedy growth, then switch to a directness-focused manual ordering—may be a practical design recipe; the paper mentions this switch as a possible improvement but does not implement it.","The buffer size of 1.5 link lengths is a single modeling choice, and the paper shows the number of links needed for full coverage swings from 20% to 100% of links as buffer size changes; re-running the strategy ranking across buffer sizes would test whether the ordering of strategies is an artifact of this choice.","Because directness-greedy growth is computationally expensive and still falls into local traps, the paper's result that cheap centrality heuristics beat directness-greedy suggests surrogate metrics could capture most of the benefit at far lower computational cost."],"forward_implications":["If the central claim holds, a planner with a fixed final network should prefer a global, hand-designed or hierarchical link order over per-step greedy optimization: it yields higher directness and coverage throughout construction, not just at the end.","Coverage-greedy orders should be used when reaching full geographic coverage early is the priority, but they should be followed by a directness-oriented phase, because after full coverage they behave like random choice and sacrifice directness.","Directness-greedy orders will underperform on directness itself on grid-like networks because they are trapped by local optima; centrality-based greedy strategies are a better directness heuristic, though they sacrifice coverage.","Additive (forwards) growth should be preferred to subtractive (backwards) growth on these stylized networks, since no subtractive strategy reaches the Pareto front except centrality-based ones—a caution against recent subtractive design methods.","Random construction, which can arise from fragmented local decision-making in real cities, is far from Pareto efficient: brute-forcing random orders would need on the order of 500 million trials to get close to the front."],"supporting_citations":[{"why":"Supplies the coverage and directness operationalization, the random-growth model for bicycle networks, and the application context that this paper's connected-growth study extends and constrains.","marker":"Szell et al. (2022)"},{"why":"Defines the directness standard (D ≥ 0.79) and the bicycle-network design requirements that motivate choosing directness and coverage as the two primary performance metrics.","marker":"CROW (2016)"},{"why":"Pioneers the subtractive, demand-driven design approach whose backwards-percolation strategy this paper's additive-versus-subtractive comparison directly addresses.","marker":"Steinacker et al. (2022)"},{"why":"Another recent subtractive method for network-wide road space reallocation used as a comparison point for the paper's finding that additive growth generally outperforms subtractive growth.","marker":"Ballo et al. (2024)"},{"why":"Recent network-science-based bicycle network design work that supports the paper's claim that such approximate formulations are justifiable despite their simplifications.","marker":"Steinacker et al. (2025)"},{"why":"A recent subtractive bike-network planning method that the paper contrasts with additive growth when arguing that additive strategies outperform subtractive ones.","marker":"Wiedemann et al. (2025)"},{"why":"Provides the spatial-network background and the closeness and betweenness centrality concepts that define the greedy centrality-based strategies.","marker":"Barthelemy (2022)"},{"why":"Justifies the planar-network assumption by showing that real street networks are mostly planar, the class of networks studied here.","marker":"Boeing (2020)"}],"fun_headline_variants":["Manual link ordering beats greedy on directness and coverage","Growth order decides network usefulness before completion","Random network growth sits six standard deviations from Pareto","Greedy directness and coverage trade off in network growth","Human-designed growth orders outperform greedy and random"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The ranking depends on the claim that the hand-designed orders C and D are globally optimal upper bounds for coverage and directness on the regular grid; the paper asserts this optimality without proof, and for the other test networks it makes no global-optimality claim at all.","fun_headline_variants_meta":{"raw":{"variants":["Manual link ordering beats greedy on directness and coverage","Growth order decides network usefulness before completion","Random network growth sits six standard deviations from Pareto","Greedy directness and coverage trade off in network growth","Human-designed growth orders outperform greedy and random"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000255,"raw_usage":{"total_tokens":1583,"prompt_tokens":968,"completion_tokens":615,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":543}},"tokens_in":584,"tokens_out":615,"duration_ms":7200,"temperature":1.0,"reasoning_tokens":543,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:31:52.502617+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For the 10×10 grid, solve the link-ordering problem exactly with an integer program or branch-and-bound search that maximizes the area under the curve of coverage or directness over all permutations of the 180 links, and compare those exact optima with the AUC values reported for manual orders C and D; if either exact optimum differs from the reported manual value, the claim that manual orders are upper bounds and therefore beat greedy strategies would be falsified.","supporting_citations":[{"cited_title":", Mimar , S","cited_arxiv_id":null,"evidence_quote":"Supplies the coverage and directness operationalization, the random-growth model for bicycle networks, and the application context that this paper's connected-growth study extends and constrains."},{"cited_title":"CROW (2016) botherref","cited_arxiv_id":null,"evidence_quote":"Defines the directness standard (D ≥ 0.79) and the bicycle-network design requirements that motivate choosing directness and coverage as the two primary performance metrics."},{"cited_title":", Storch , D.-M","cited_arxiv_id":null,"evidence_quote":"Pioneers the subtractive, demand-driven design approach whose backwards-percolation strategy this paper's additive-versus-subtractive comparison directly addresses."},{"cited_title":", Raubal , M","cited_arxiv_id":null,"evidence_quote":"Another recent subtractive method for network-wide road space reallocation used as a comparison point for the paper's finding that additive growth generally outperforms subtractive growth."},{"cited_title":", Paulsen , M","cited_arxiv_id":null,"evidence_quote":"Recent network-science-based bicycle network design work that supports the paper's claim that such approximate formulations are justifiable despite their simplifications."},{"cited_title":": Spatial Networks: a Complete Introduction : from Graph Theory and Statistical Physics to Real-world Applications","cited_arxiv_id":null,"evidence_quote":"Provides the spatial-network background and the closeness and betweenness centrality concepts that define the greedy centrality-based strategies."},{"cited_title":": Planarity and street network representation in urban form analysis","cited_arxiv_id":null,"evidence_quote":"Justifies the planar-network assumption by showing that real street networks are mostly planar, the class of networks studied here."}],"review_version":1}