{"id":"496da676-a9b5-4714-b85d-89deb0df1a1e","arxiv_id":"2507.00335","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A recursive graph-merging algorithm produces networks with extreme topological properties that their degree sequences do not predict.","lead":"This paper introduces a graph generator, inspired by assembly theory, that repeatedly merges randomly chosen pairs of graphs. The resulting networks have unusually high clustering, betweenness, and diameter compared to random graphs with the same degree sequence, a diversity the authors trace to historical path dependence.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Fig. 4 control randomizes connectivity without preserving degree sequences, so the claimed history effect may be a degree-sequence artifact; a degree-preserving rewiring control is required.","rationale":"The reader's weakest-assumption analysis and my own independent reading converge on the same load-bearing concern: the Fig. 4 control does not isolate historical contingency from degree-sequence effects. This is the most critical point because it directly undermines the paper's causal conclusion that 'historical contingencies steer the topology.' The descriptive result in Fig. 2—that assembled graphs are extreme relative to configuration-model graphs with matched degree sequences—is robust to this concern and remains a valuable contribution. However, without a degree-preserving control, the paper has not demonstrated that the generative history, rather than the degree sequences it produces, drives the topological anomalies. The proposed test settles this by comparing original graphs to a control that randomizes connectivity while holding degree sequences fixed. The paper's other issues (e.g., disconnected-component handling, missing statistical test details) are real but secondary; they affect reproducibility and precision, not the central causal inference. Since the reader already assigned CONDITIONAL based on this concern, my read does not change the verdict; it remains CONDITIONAL pending the degree-preserving control.","tokens_in":7104,"tokens_out":6695,"duration_ms":73592,"concrete_test":"Re-run the Fig. 4 control using a configuration-model rewiring that exactly preserves the degree sequence of each chosen L and R before merging (and preserves the chosen vertex pairs), rather than randomizing connectivity with only node/edge counts held fixed. For each of the 1000 trajectories, generate the original final graph and the degree-preserving control final graph, and compare the distributions of clustering coefficient, mean betweenness, diameter, and algebraic connectivity. If the original and degree-preserving control distributions remain significantly distinct, the historical-contingency claim survives; if they converge (e.g., the original's statistics fall within the control's spread or the Kolmogorov–Smirnov test becomes non-significant), the Fig. 4 difference is a degree-sequence artifact and the causal claim would need to be weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim that historical contingencies steer topology rests on the Fig. 4 control experiment. In that control, after L and R are chosen, 'their connectivity is randomized, such that the number of nodes and edges are preserved and the graph is still connected, but otherwise random' (Results, paragraph before Fig. 4). This manipulation changes the degree sequences of L and R, and hence of the merged graph. The original assembled graphs have degree sequences that are themselves products of the historical reuse process; by randomizing connectivity without holding degree sequences fixed, the control conflates two effects: (i) the effect of the specific historical sequence of graph reuse, and (ii) the effect of the degree sequence that such history produces. Since degree sequence alone is known to strongly constrain clustering, betweenness, diameter, and algebraic connectivity (as the paper's own Fig. 2 demonstrates against configuration-model controls), the observed differences between original and control distributions in Fig. 4 could be entirely explained by the control's altered degree sequences. The Fig. 4 caption's conclusion—'historical dependence of graph reuse is partially responsible for the observed topological properties'—is therefore not uniquely supported. The paper notes that the control 'controls for the mean degree,' but mean degree is a drastically weaker constraint than the full degree sequence. This is a correctness risk for the causal claim, not for the descriptive finding that assembled graphs are atypical relative to their own degree sequences (Fig. 2).","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a random graph assembly process inspired by assembly theory. Starting from a multiset containing a two-node path and a three-node path, each iteration selects two graphs (with a bias parameter p for choosing from the largest graphs), merges a random number M of vertex pairs, and adds the resulting graph to the multiset. The authors generate trajectories of 25 iterations, focus on the largest graph in each trajectory, and compute global clustering coefficient, mean betweenness centrality, diameter, and algebraic connectivity. They compare these statistics to those of configuration-model graphs with identical degree sequences and report that assembled graphs frequently fall in extreme tails of the control distribution (Fig. 2). They then introduce two controls: a history-loss resampling (Fig. 3) and a connectivity-randomized control that preserves node count, edge count, and connectedness (Fig. 4). Based on these controls they conclude that historical contingencies during the generative process steer the topology of assembled graphs.","tokens_in":7310,"tokens_out":6404,"duration_ms":76629,"significance":"If the main claims hold, the paper contributes a new random-graph ensemble that is structurally distinct from Erdős-Rényi, Watts-Strogatz, Barabási-Albert, and Kronecker models, and it provides a concrete algorithmic bridge between assembly theory and graph sampling. The pseudocode and public code are strengths, and the Z-score protocol in Fig. 2 is a reproducible, falsifiable way to quantify atypicality relative to degree-sequence controls. The potential applications to chemical-space and material sampling make the questions worth pursuing. However, the causal claim about historical contingency is currently supported by a control that does not isolate history from degree-sequence effects, and the degree-sequence comparison has an asymmetry that needs correction; these issues affect load-bearing parts of the paper's central claims.","major_comments":[{"comment":"The connectivity-randomized control preserves only the number of nodes, the number of edges, and connectedness of the two selected graphs; it does not preserve their degree sequences, and therefore does not preserve the degree sequence of the merged graph. Since Fig. 2 demonstrates that degree sequence alone strongly constrains clustering, betweenness, diameter, and algebraic connectivity, the distributional differences between original and control in Fig. 4 could be fully explained by the control's altered degree sequences rather than by 'historical dependence of graph reuse.' The caption's claim that historical dependence is 'partially responsible' for the observed topological properties is therefore not uniquely supported by this experiment. A degree-preserving rewiring control (for example, double-edge swaps on L and R after selection, or a configuration-model resampling restricted to connected graphs) is necessary to separate the historical-reuse effect from the degree-sequence effect.","section":"Results, paragraph before Fig. 4; Fig. 4 caption"},{"comment":"The paper compares assembled graphs, which are connected by construction, to configuration-model graphs with identical degree sequences, but when the randomized graphs are disconnected it computes the statistics on the largest connected component. The largest connected component of a disconnected configuration-model graph does not have the degree sequence of the assembled graph, so the comparison is no longer between graphs with identical degree sequences. This asymmetry can bias all four Z-scores reported in Fig. 2 and undermines the quantitative support for the claim that assembled graphs are atypical relative to graphs with identical degree sequences. The atypicality claim should be evaluated against connected configuration-model graphs (for example, by conditioning on connectedness or by using a connected realization algorithm), or by treating disconnected graphs symmetrically.","section":"Results, Fig. 1C and Fig. 2; text on disconnected randomizations"},{"comment":"The history-loss experiment in Fig. 3 documents variation in mean degree as a function of how many final steps are resampled, but the paper's central topological claim concerns clustering, betweenness, diameter, and algebraic connectivity. Mean degree alone is a much weaker proxy for those properties, so Fig. 3 does not directly demonstrate that the exceptional topology of Fig. 2 depends on historical contingency. Since Fig. 4 is the only direct evidence connecting these topological properties to history, and that control is confounded as described above, the conclusion that 'the algorithm's historically contingent features drive these key properties' is currently under-supported. Directly tracking the four topological statistics in history-loss ensembles, or coupling a degree-preserving control with trajectory-level analysis, would close this gap.","section":"Results, Fig. 3 and Conclusion"}],"minor_comments":[{"comment":"The caption states that the cones indicate the maximum and minimum mean degree observed at each iteration, while the text says the intervals are the inter-quantile range between 1% and 99%; these descriptions should be reconciled.","section":"Fig. 3 caption"},{"comment":"The text reports that the distributions of original and control graphs are 'all statistically significant,' but no test name, sample sizes, or p-values are given; please specify the significance testing procedure.","section":"Fig. 4"},{"comment":"The appendix says the number of merges is 'at minimum the smaller of the two graphs' vertex counts,' but the algorithm caps the selected M at that smaller count; the wording should be corrected to 'at most.'","section":"Appendix B"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the journal's scope as a generative-model paper in physics of social and complex systems. The central causal claim about historical contingency rests on the Fig. 4 control, which is confounded by degree-sequence changes; this is fixable with a degree-preserving rewiring control. The Fig. 2 comparison also needs to be made symmetric with respect to connectedness. I see no circularity or fabrication concerns, and the code availability is a clear strength. I recommend major revision rather than rejection because the core generative model and the atypicality observation are likely salvageable with additional controls."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Cole, quick take on 2507.00335. The paper does something real: it defines a new generative model for graphs based on recursive merging, ships code, and gives a clean Z-score analysis showing that the assembled graphs are often extreme relative to configuration-model controls with the same degree sequence. The algorithm is a genuine addition to the random graph toolkit, and the descriptive result that assembled graphs are atypical relative to their degree sequences is the core contribution. That part looks solid, at least as far as the analysis goes. The history-loss experiment in Fig 3 is also a nice touch.\n\nThe soft spot is the Fig 4 control, and it's exactly the one flagged in the stress-test. The control randomizes connectivity while preserving only node and edge counts, not the degree sequence. The assembled graphs' degree sequences are themselves shaped by the history of reuse; randomizing connectivity erases that degree structure. So the difference between original and control could be explained by degree effects alone, not by 'historical contingencies' in the sense the paper means. The caption's conclusion that historical dependence is 'partially responsible' is not uniquely supported. This needs a degree-preserving rewiring control, or at least an explicit check that the degree sequences of original and control graphs are matched.\n\nTwo smaller issues. First, disconnected configuration-model graphs are compared via their largest connected components, while assembled graphs are connected. That can distort the degree sequence being conditioned on and bias the Z-scores. Second, the paper claims the Fig 4 distributions are 'all statistically significant' but gives no test details, which is unverifiable as written.\n\nNone of this kills the paper. The descriptive finding is plausible and useful; the causal story is the weak link and is fixable. This deserves a serious referee, and I'd expect a major revision focused on the Fig 4 control. I'd bring it to our reading group and would cite the descriptive result if I worked on generative graph models.","headline":"New generative model, solid descriptive result, but the historical-contingency claim rests on a control that fails to preserve degree sequences.","tokens_in":7849,"tokens_out":2450,"would_cite":true,"duration_ms":26101,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["89.75.Fb","89.75.Hc"],"model":"deepseek-v4-flash","headline":"A new random graph model, inspired by assembly theory, samples networks whose topological statistics are extreme outliers relative to their degree sequences.","keywords":["random graph model","assembly theory","historical contingency","degree sequence","configuration model","network topology","generative model","drug discovery"],"falsifier":"Recompute the Fig. 4 comparison with a control that preserves the degree sequence of each merged pair—for instance, by rewiring with configuration-model swaps at every step. If the original and degree-preserving control distributions still differ significantly, historical contingency is confirmed; if they match, the effect is explained by degree structure instead.","tokens_in":6880,"feed_emoji":"🕸️","tokens_out":6158,"duration_ms":60644,"temperature":0.7,"pith_summary":"The paper introduces a random graph model inspired by assembly theory: starting from a multiset of two small path graphs, at each step two graphs are chosen (one possibly biased toward the largest), some of their nodes are merged, and the result is added back to the multiset. The authors show that after 25 iterations the largest assembled graphs frequently land in the extreme tails of the distribution of four topological statistics—clustering coefficient, mean betweenness, diameter, and algebraic connectivity—computed from configuration-model random graphs with identical degree sequences. They further argue that this atypicality is steered by historical contingency: the specific sequence of merges within a trajectory determines which part of graph space is reached, and randomizing connectivity while preserving only node and edge counts eliminates the effect. The result matters because it provides a generative sampler that explores unusual regions of graph space, with stated applications in drug discovery and materials science.","feed_headline":"Merging graph pairs yields networks that defy their degree sequence","feed_subtitle":"A new assembly-based sampler often lands in the extreme tails of clustering, diameter, and centrality.","key_machinery":"The central object is the graph assembly algorithm itself, a stochastic recursive merging process. It initializes a multiset with a path graph on two nodes and a path graph on three nodes. At each iteration it selects a left graph L and right graph R; L is chosen uniformly from the largest graphs with probability p, otherwise uniformly from the whole multiset, while R is always uniform. It then merges M randomly chosen node pairs, one node from each graph, combining their neighborhoods, removing parallel edges and self-loops, and returns the resulting simple graph to the multiset. The algorithm's historical contingency comes from the reuse of previously assembled graphs: once a graph is generated it becomes a possible component in future merges, so the particular trajectory of selections and merges is encoded in the final topology.","core_discovery":"On the paper's own terms, the central discovery is that a recursively generative process—repeatedly merging pairs of graphs from an ever-growing multiset—produces an ensemble whose members are often statistically exceptional relative to their degree sequences. Quantitatively, for 1000 assembled graphs compared against 1000 configuration-model controls each, a large fraction of assembled graphs had absolute Z-scores above 2 and above 4 for at least one topological measure, meaning that standard random-graph nulls would rarely produce such structures. The authors identify the cause as historical contingency: because early graphs persist in the multiset and can be reused, random fluctuations early in a trajectory become amplified and steer later graphs toward distinct, extreme topologies. They support this by showing that if the connectivity of the two graphs being merged is randomized at each step (preserving node and edge counts and connectedness), the distinct topological properties largely disappear.","pith_inferences":["A strict test of the historical-contingency claim would need a control that preserves each merged graph's degree sequence while scrambling only the order of operations; the paper's Fig. 4 control also removes degree structure, so the causal role of history is not fully isolated.","The same algorithm with different initial graph multisets (e.g., triangles or stars) may generate families of ensembles with different extreme-tail behavior, turning the seed set into an additional design parameter.","The extreme atypicality of assembled graphs suggests that such graphs could serve as challenging nulls or anomaly probes for network science: any statistical test tuned to configuration-model expectations would classify them as non-random."],"forward_implications":["The assembly sampler can generate graphs that are extreme outliers relative to their own degree sequences, offering a complementary null or generative model to Erdős–Rényi and configuration-model ensembles.","Tuning the bias parameter p and the merge range M controls both graph size and the diversity of accessible topologies, so the ensemble can be tailored to explore specific regions of graph space.","Because history steers the outcome, two trajectories that start identically and use the same parameters can converge to very different graphs, implying that trajectory-level (not just parameter-level) information controls the sampling.","The graph-level formalization extends earlier assembly-theory work on molecular assembly trees, so the same procedure can be applied to construct molecules or materials with rare, drug-like properties."],"supporting_citations":[{"why":"Defines assembly theory and the assembly index that motivate the constructive reuse of subgraphs.","marker":"[15]"},{"why":"Formalizes assembly spaces, the pathway structure the graph algorithm instantiates.","marker":"[16]"},{"why":"The prior molecular assembly-tree approach that the graph algorithm generalizes from molecules to arbitrary graphs.","marker":"[21]"},{"why":"Defines the Erdős–Rényi random graph model used as a baseline ensemble.","marker":"[7]"},{"why":"Provides the standard graph measures (clustering, betweenness, diameter) and the configuration-model context.","marker":"[11]"},{"why":"Defines algebraic connectivity, one of the four topological statistics used to characterize assembled graphs.","marker":"[24]"},{"why":"Introduces Kronecker graphs, a recursive generative model the authors compare conceptually with their own.","marker":"[13]"}],"fun_headline_variants":["Merge history steers graph topology beyond degree constraints","Assembly-inspired merges produce statistically exceptional graphs","Contingent merges push graphs to extreme topological tails","Historical chance in assembly yields graph topologies beyond degree"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the difference between the original algorithm and the Fig. 4 control is caused by historical contingency; but the control randomizes each merged graph's connectivity while preserving only node and edge counts, so it also erases degree-sequence structure, and the measured difference could be due to that loss rather than to the order of assembly steps.","fun_headline_variants_meta":{"raw":{"variants":["Merge history steers graph topology beyond degree constraints","Assembly-inspired merges produce statistically exceptional graphs","Contingent merges push graphs to extreme topological tails","Historical chance in assembly yields graph topologies beyond degree"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000606,"raw_usage":{"total_tokens":2773,"prompt_tokens":842,"completion_tokens":1931,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":458,"completion_tokens_details":{"reasoning_tokens":1870}},"tokens_in":458,"tokens_out":1931,"duration_ms":17405,"temperature":1.0,"reasoning_tokens":1870,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:17:32.161695+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the Fig. 4 comparison with a control that preserves the degree sequence of each merged pair—for instance, by rewiring with configuration-model swaps at every step. If the original and degree-preserving control distributions still differ significantly, historical contingency is confirmed; if they match, the effect is explained by degree structure instead.","supporting_citations":[{"cited_title":"Sharma, D","cited_arxiv_id":null,"evidence_quote":"Defines assembly theory and the assembly index that motivate the constructive reuse of subgraphs."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Formalizes assembly spaces, the pathway structure the graph algorithm instantiates."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The prior molecular assembly-tree approach that the graph algorithm generalizes from molecules to arbitrary graphs."},{"cited_title":"Erdos, A","cited_arxiv_id":null,"evidence_quote":"Defines the Erdős–Rényi random graph model used as a baseline ensemble."},{"cited_title":"Fiedler, Algebraic connectivity of graphs, Czechoslo- vak mathematical journal23, 298 (1973)","cited_arxiv_id":null,"evidence_quote":"Defines algebraic connectivity, one of the four topological statistics used to characterize assembled graphs."},{"cited_title":"Leskovec, D","cited_arxiv_id":null,"evidence_quote":"Introduces Kronecker graphs, a recursive generative model the authors compare conceptually with their own."}],"review_version":1}