{"id":"5b9b2154-6855-4842-b0b2-1a15191aaf7a","arxiv_id":"2512.02362","paper_version":3,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":9,"one_line_summary":"A four-step algorithm reconstructs millions-firm production networks from public sector flows and firm-size distributions, but key validation claims appear only in the abstract.","lead":"This paper builds synthetic national firm-to-firm production networks from sector input-output tables and firm-size bins, using a gravity model, random link draws, and an optimization that sets link weights. It claims to reproduce observed degree patterns and to enable firm-failure risk experiments without proprietary data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The weighting QP constrains sector size totals but never the bilateral sector-pair IO flows; the central 'preserves sectoral flows' claim is not implied by the algorithm.","rationale":"The reader's weakest assumption -- the arbitrary $1B cap on the largest SBA size bin -- targets the secondary claim of reproducing a heavy-tailed degree distribution. The first clause of the central claim, preservation of sectoral flows, is more fundamental and fails for a different reason: the optimization in Section 3.4 never imposes constraints on bilateral sector-pair flows. The firm-level and sector-total constraints can all be satisfied while the 24x24 flow matrix differs arbitrarily from the BEA IO table. The paper's own language ('sectoral flows of the economy upon the aggregation of firm-to-firm flows') invites the natural reading that the IO matrix, not merely sector size totals, should be preserved. A simple aggregation-and-compare test would settle this. If the test fails, the paper's headline contribution is not established regardless of degree-tail behavior; if it passes, the concern is resolved. Since the reader already recommends REJECT and this separate concern reinforces that verdict, I leave the verdict unchanged.","tokens_in":29090,"tokens_out":11790,"duration_ms":131277,"concrete_test":"Reconstruct the US network with the same pipeline; form the 24x24 aggregated flow matrix F_{kl} = sum_{i in sector k, j in sector l} m_i w_{ij}. Rescale F to the same total transactions as the BEA IO table used in Section 5.1 and compute the normalized L1 error ||F - IO||_1 / ||IO||_1 plus per-cell relative errors. If the normalized error is far above the 10% sector-total tolerance (e.g. >30%) while the sector totals are within 10%, the algorithm preserves sector size totals but not sectoral flows, and the central claim must be weakened or the QP amended with IO-pair constraints.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.4's minimum-energy QP minimizes ||W||_F^2 subject to row-stochasticity, support/bounds, firm-level one-step balance ((W^T m)_j - m_j)^2 <= delta^2 m_j^2, sector totals |hat{s}_ell(W)-s_ell| <= epsilon s_ell, and self-loop caps. It does not constrain the 24x24 input-output matrix IO_{kl}. The gravity model in Section 3.1 only shapes the binary support through link probabilities; the weights are otherwise free to move money across seller sectors while satisfying column/sector totals. Hence the aggregated matrix F_{kl} = sum_{i in k, j in l} m_i w_{ij} need not reproduce the BEA IO flows. The Introduction's central assertion -- that aggregating firm-to-firm flows preserves sectoral flows of the economy -- therefore does not follow from the optimization as written. With millions of firms and only firm-level plus 24 sector constraints, the weight matrix is massively underdetermined, so this is a structural gap, not a minor tolerance issue.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a four-step algorithm to reconstruct large, weighted firm-to-firm production networks from publicly available sectoral input-output tables and firm-size-by-sector distributions. It estimates link probabilities with a logistic-gravity model, draws a Bernoulli backbone, enforces irreducibility/aperiodicity through a Markov closure, and assigns weights via a convex minimum-energy QP with firm-level and sectoral balance constraints. The authors reconstruct a US network of several million firms and tens/hundreds of millions of links, compare summary statistics with an observed US sample, and claim that the algorithm preserves sectoral flows upon aggregation and reproduces known granular production-network properties. The abstract also promises a Japanese heavy-tail validation and firm-failure experiments.","tokens_in":29466,"tokens_out":6353,"duration_ms":69351,"significance":"The open-source code release, explicit Markov-regularity step, and the sparsity/binning complexity analysis are real strengths; the concentration remarks in Appendix A provide formal support for working with a single Bernoulli draw. If the sectoral-flow preservation property were actually enforced and the promised validations supplied, this would be a valuable tool for calibrated shock-propagation studies at national scale. However, the manuscript as submitted does not support its central claim as written, and several headline validations announced in the abstract are absent from the body, so the significance cannot yet be assessed at face value.","major_comments":[{"comment":"The QP in §3.4 constrains only sector-level totals ĥ{s}_\\ell(W), never the bilateral IO flows IO_{kl}. The gravity model in §3.1 shapes only the binary support through p_{ij}; the weights are otherwise free to distribute money across seller sectors subject to row sums and aggregate sector constraints. Therefore the aggregated matrix F_{kl} = \\sum_{i\\in k, j\\in l} m_i w_{ij} need not reproduce the 24x24 BEA table, and the central assertion in the Introduction and Concluding Remarks that the algorithm 'preserves sectoral flows' is not implied by the optimization. This is a structural gap, not a tolerance issue.","section":"§3.4; Introduction"},{"comment":"The abstract promises that the Japanese reconstruction, built without link data, reproduces Japan's heavy-tailed degree regime, and that firm-failure experiments show that no simple firm attribute proxies aggregate losses. The body contains no Japan section and no firm-failure simulations; the only validation is the US reconstruction in §5, followed by the factory-level extension in §6. These claims must either be backed by the reported experiments or removed from the abstract.","section":"Abstract; §§5–7"},{"comment":"Firm sizes are generated by uniform draws within SBA bins with an arbitrary $1B cap on the largest bin. Section 5.3 attributes the steep tail descent in Figure 7 to this cap and explicitly states that the degree distribution would differ if a larger endpoint were used. Since the heavy-tailed degree regime is a headline validation, this untested distributional assumption is load-bearing. The paper should provide sensitivity analyses with alternative caps/within-bin distributions and replace the visual 'not a bad fit' with a formal tail test or power-law fit.","section":"§5.2.1, §5.3, Figure 7"},{"comment":"The numerical claims are inconsistent: the abstract reports about 6.5 million firms and 340 million links reconstructed in roughly four hours, while §5.2.2 reports 5.4 million firms and 1.25×10^8 connections, and §5.3 says 'almost 1.3×10^8' directed links. Section 5.2.2 also reports 48 hours of machine time for parameter estimation. These numbers must be reconciled.","section":"Abstract; §5.2.2, §5.3"},{"comment":"Table 2 reports O(N_F) serial complexity for Step 3, but §4 states that the Markov-closure problem is a mixed-integer quadratic program that is NP-hard in general and has no polynomial worst-case bound. The claim that the binned pipeline is 'essentially linear' is therefore not justified for the complete algorithm. The complexity claims need to be qualified with explicit assumptions under which the MILP stage solves in linear time, or the table/text must be revised.","section":"§4, Table 2"},{"comment":"The closeness of the stationary money vector to the empirical size vector is not an independent validation: the firm-level constraints ((W^T m - m)/m)^2 ≤ δ^2 impose one-step balance by construction, and Remark 4's L1 bound is a direct consequence. The paper should present this as a design guarantee built into the algorithm, not as an ex-post empirical check, and should not count it as separate evidence that reconstructed networks match observed stationary firm sizes.","section":"§3.4, Remark 4"}],"minor_comments":[{"comment":"The abstract ends mid-sentence with 'and a graphical'; either complete this or delete it.","section":"Abstract"},{"comment":"Typo: 'paramerter' should be 'parameter'.","section":"Figure 3 caption"},{"comment":"The reference to 'Bacilieri and Austudillo-Estevez' contains a misspelling; it should be 'Astudillo-Estevez'.","section":"References"},{"comment":"The symbol I_{kl} is used both for the row-stochastic IO matrix (Section 3.1) and for a 0/1 sectoral incidence matrix (Section 3.3). These are different objects; use distinct notation to avoid ambiguity.","section":"§3.1, §3.3"},{"comment":"The range reported for λ is very broad ('10^{-2} to 2'); with 173 estimated sector multipliers, a full table or a more detailed summary (e.g., quantiles) would help reproducibility.","section":"Table 3"},{"comment":"The comparison with the S&P Capital IQ sample would benefit from a fuller description of that sample's industry and size composition, since the reconstructed network is built to match sectoral aggregates and the comparison sample is only 78k nodes.","section":"§5.3, Table 5"}],"recommendation":"major_revision","confidential_remarks":"The missing Japan section and firm-failure experiments are the most serious issue: the abstract currently promises results that the body does not deliver. I would ask the authors to either add those results or substantially revise the claims. The sectoral-flow preservation gap in §3.4 is fixable in principle by adding bilateral IO-pair constraints to the weighting QP, but until that is done the paper's main advertised property is unsupported. If the authors are unwilling to add such constraints or to provide the missing validations, rejection would be warranted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know: this is a serious attempt at a hard problem, but the submitted version overclaims in ways that matter. The four-step pipeline—gravity-based link probabilities, Bernoulli backbone, Markov regularization, minimum-energy weighting—is genuinely new as a combination, and running it at 5+ million firms is a real step beyond prior work like Ialongo et al.'s 10^5-firm reconstructions. The writing is clear, the authors acknowledge the arbitrary $1B firm-size cap and its effect on the degree tail, and the bootstrap stability check on the gravity parameters is a nice touch. Credit where due: the method is coherent and the computational scaling story is plausible.\n\nNow the soft spots. The stress-test concern is correct and lands on the paper's central promise. The weighting QP in Section 3.4 constrains firm-level one-step balances and sector size totals, but never constrains the bilateral sector-pair IO matrix. The gravity model shapes the binary support using IO flows, but the weights are free to move money across seller sectors as long as column totals and firm balances are within tolerance. So the introduction's claim that aggregating firm-to-firm flows 'preserves sectoral flows' does not follow from the optimization as written. The paper never validates that the aggregated F_kl matches the BEA table. That is a structural gap, not a tolerance issue.\n\nThe abstract promises two things the body does not deliver: the Japanese heavy-tailed degree reproduction and the firm-failure systemic-risk experiments. Neither appears anywhere in the full text. The reported scale is also internally inconsistent—the abstract says 6.5M firms and 340M links, the body says 5.4M and 1.3e8, and the stated density of 2.4e-5 does not match the reported link count for a network of that size. The code repository is announced but explicitly not yet integrated or usable. The stationarity bound in Remark 4 depends on the spectral gap gamma, which is never quantified, and Remark 7 assumes the existence of an exactly money-preserving matrix on the support—a strong assumption that is not tested.\n\nNone of this kills the underlying idea. The pipeline is sensible and the Markov regularity contribution is a useful innovation. But the paper as submitted is not ready. It needs a serious referee, not a desk reject, because the method is novel and potentially important. The right path is peer review with the expectation of major revision: add or remove the missing validations, constrain or soften the sectoral-flow claim, fix the numeric inconsistencies, and ship usable code. Who should read it: economists and network scientists working on granular shock propagation, especially those needing synthetic firm-level networks. I'd bring it to a reading group as a case study in claims versus constraints, but I would not cite it in its current form.","headline":"A serious, large-scale reconstruction pipeline whose headline claim—preserving sectoral IO flows—is not enforced by its constraints, with several promised validations missing from the submitted text.","tokens_in":29900,"tokens_out":3098,"would_cite":false,"duration_ms":36212,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper tries to establish that a country's firm-to-firm production network can be reconstructed from public sectoral-flow and firm-size data well enough to preserve aggregate flows and reproduce the heavy-tailed, low-reciprocity topolog","keywords":["production networks","network reconstruction","input-output tables","firm size distribution","gravity model","Markov chain","minimum-energy weighting","systemic risk"],"falsifier":"Compare the reconstructed Japanese degree distribution to the observed Japanese network's degree distribution across the full support; if the reconstructed distribution falls outside the concentration bounds implied by the Bernoulli ensemble, the paper's claim that the algorithm reproduces widely reported granular properties would be falsified.","tokens_in":28945,"feed_emoji":"🏭","tokens_out":7134,"duration_ms":73124,"temperature":0.7,"pith_summary":"This paper argues that a country's firm-to-firm production network can be reconstructed from two public objects—sectoral input-output flows and the distribution of firm sizes by sector—well enough to reproduce the properties that matter for aggregate shock propagation. The proposed algorithm draws a binary buyer-seller backbone from a sector-aware gravity model, closes it to make the network primitive (irreducible and aperiodic), and then assigns weights by a minimum-energy program that keeps sectoral flows and stationary firm sizes close to the data. The payoff, if the claim is right, is that researchers can run granular impulse-response and systemic-risk simulations on synthetic country-scale networks (millions of firms) without access to confidential firm-level data. The paper demonstrates the pipeline on the United States and several other countries, and reports that the Japanese reconstruction reproduces the heavy-tailed degree regime observed in the true network.","feed_headline":"Four steps rebuild a country's firm-to-firm network from public data","feed_subtitle":"It preserves sectoral flows and reproduces heavy-tailed connectivity, enabling shock simulations without proprietary data.","key_machinery":"The load-bearing object is the sector-aware logistic-gravity link-probability matrix P: the probability that firm i buys from firm j is x/(1+x) with intensity x = z times lambda_kl times S_kl^kappa times (m_i m_j)^alpha, where S_kl scales the sector-to-sector flow and m_i is normalized firm size. That matrix drives the Bernoulli backbone; a Markov-closure step (adding strongly-connected-component bridges and self-loops) makes the graph irreducible and aperiodic; and a strictly convex quadratic program minimizes total squared edge weights subject to row-stochasticity, support, firm-level and sectoral consistency, and self-loop caps, yielding a unique weight matrix whose stationary distributio","core_discovery":"The central claim is that the unobserved weighted network of buyer-seller relations among every firm in an economy can be reconstructed from aggregate public data while preserving sectoral flows upon aggregation and faithfully reproducing widely reported granular-network properties. On the paper's own terms, the reconstruction is not a guess at individual links but a statistically faithful ensemble: a logistic-gravity probability model converts firm sizes and sectoral flows into link probabilities; Bernoulli draws generate a binary backbone; a Markov-closure step adds a small number of cross-component edges and self-loops so the network has a unique stationary distribution; and a convex mini","pith_inferences":["Beyond the paper's claims, a reader should treat the reconstructed link set as an ensemble representation rather than as the true buyer-seller relations; firm-specific conclusions would require validation this method does not provide.","Beyond the paper's claims, the steep tail drop near 10^4 connections is, by the authors' own admission, a consequence of the $1 billion cap on the largest size bin, so the true heavy tail is likely longer than the reported reconstruction shows.","Beyond the paper's claims, applying the same pipeline to world input-output tables and country-level size distributions would yield a global firm-level network, enabling cross-border climate and disaster simulations.","The algorithm treats buyers and sellers symmetrically, yet the reconstructed networks show heavier customer tails than supplier tails; testing whether this asymmetry persists across countries would clarify whether it emerges from the sectoral flow structure itself."],"forward_implications":["Because the reconstructed network is irreducible and aperiodic, it has a unique stationary distribution, so impulse-response analyses after firm-level shocks are well-defined and do not depend on arbitrary initial conditions.","National-scale network reconstruction becomes feasible as a routine computation: with sector-wise size binning the pipeline scales essentially linearly in the number of firms, and the authors report building a 5.4-million-firm US network in hours on one workstation.","Failure experiments on the reconstructed US network indicate that neither firm size, degree, nor sectoral position predicts the aggregate loss from a firm's failure, so studying systemic risk requires the full weighted network rather than summary attributes.","Countries that publish only input-output tables and size counts can obtain synthetic firm-level networks for shock-propagation research; the Japanese case offers an out-of-sample check, since no Japanese link data were used in the reconstruction.","Sectoral flows are preserved upon aggregation, so the reconstructed networks can be embedded in existing macroeconomic input-output frameworks without breaking aggregate accounting identities."],"fun_headline_variants":["Four public-data steps rebuild a nation's firm network","Reconstructing 340M firm links from just sectoral tables","Public IO tables spawn full buyer-seller networks","Algorithm turns aggregate flows into firm-level maps","No proprietary data needed to map production networks"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the true distribution of firm sizes inside each published size bin is well approximated by uniform sampling with a $1 billion cap on the largest bin; the paper itself notes the degree tail would change if that cap were larger or the bin were non-uniform.","fun_headline_variants_meta":{"raw":{"variants":["Four public-data steps rebuild a nation's firm network","Reconstructing 340M firm links from just sectoral tables","Public IO tables spawn full buyer-seller networks","Algorithm turns aggregate flows into firm-level maps","No proprietary data needed to map production networks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001054,"raw_usage":{"total_tokens":4281,"prompt_tokens":783,"completion_tokens":3498,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":527,"completion_tokens_details":{"reasoning_tokens":3424}},"tokens_in":527,"tokens_out":3498,"duration_ms":28699,"temperature":1.0,"reasoning_tokens":3424,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T18:59:29.794771+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare the reconstructed Japanese degree distribution to the observed Japanese network's degree distribution across the full support; if the reconstructed distribution falls outside the concentration bounds implied by the Bernoulli ensemble, the paper's claim that the algorithm reproduces widely reported granular properties would be falsified.","supporting_citations":[],"review_version":1}