{"id":"6c4a3b88-9af3-4e82-a417-2fd5b84c60b0","arxiv_id":"2607.03971","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.5,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"OCDM recovers multistage causal DAGs by stage-constrained score-based order search plus STG-NN pruning, outperforming baselines on layered synthetic and causalAssembly data.","lead":"OCDM finds causal graphs for multistage data by first ordering variables stage-by-stage with known process knowledge, then pruning with stochastic-gated nets. This can give more accurate, stage-respecting causal maps for manufacturing and similar sequential systems than generic discovery tools.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"STG-NN pruning lacks recovery guarantees and may over-prune true edges, so the multi-metric gains on layered graphs rest on an unproven final step.","rationale":"The Reader correctly flags the additive-noise + known-stage assumptions as the weakest modeling premises and assigns a CONDITIONAL verdict. Those assumptions are necessary for the leaf-identification guarantee and for the stage-wise search to be valid, yet the paper already supplies partial robustness checks (noisy stage labels up to p=0.3, mixed-type data). The more immediate threat to the strongest claim is therefore not the generative-model assumptions but the unproven pruning step that turns the order-connected graph into the final DAG. Because the empirical superiority is demonstrated only after STG-NN pruning, and because the authors acknowledge the absence of recovery guarantees and residual over-pruning, the claim that OCDM “recovers the causal structure \\ldots more accurately” rests on an empirically convenient but theoretically unsupported final stage. The concrete parent-set TPR/FPR comparison would settle whether the gains are genuine structure recovery or merely better sparsity. This does not overturn the Reader’s CONDITIONAL verdict; it simply relocates the principal correctness risk from the modeling assumptions to the pruning procedure, keeping the overall assessment unchanged.","tokens_in":16778,"tokens_out":686,"duration_ms":6809,"concrete_test":"On the Simulation-2 graphs (w=20,k=10), compute the fraction of true parents retained after STG-NN pruning versus after CAM pruning (true-positive rate of the parent sets) and the corresponding false-positive rate; if STG-NN’s TPR is materially lower while FPR is also lower, the multi-metric gains are sparsity artifacts rather than improved structure recovery.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The strongest claim is that stage-wise DiffAN order search plus STG-NN pruning recovers multistage causal structure more accurately than baselines (Figs. 2–5). The order-search half is well-supported: topological-order divergence (D_top) is lowest for OCDM, and the leaf-identification argument (Var[∂s_i/∂x_i]=0 for leaves under Eq. 1) is standard. The load-bearing soft spot is the pruning half. Section III-D and Algorithm 1 lines 17–20 replace CAM’s group-LASSO spline selection with an STG-NN whose stochastic gates (Eq. 4) are trained by L0-relaxed regression; edges with final gate z_j=0 are discarded. No consistency or parent-set recovery theorem is given for this procedure under the nonlinear additive model. The authors themselves note (Sec. V) that both STG-NN and CAM pruning “lack theoretical guarantees for recovering the true parent set” and that “over-pruning” of true edges remains a problem. Consequently the reported AUROC/AUPRC/SHD/SID improvements could be driven by aggressive sparsity rather than faithful recovery of the layered DAG. If the gates systematically drop true parents once candidate sets become large (Simulation 2, causalAssembly), the headline claim that OCDM recovers the causal structure is only partially substantiated.","agreement_with_reader":"partial"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes OCDM, an order-based causal discovery method for multistage processes whose data-generating process is assumed to be a nonlinear additive Gaussian noise model that forms a layered DAG (Sec. III-A, Eq. 1). Stage membership of every variable is treated as known process knowledge. OCDM first runs a stage-wise leaf-node search (Algorithm 1, lines 1–14) that restricts DiffAN-style score-matching order search to the current stage, thereby guaranteeing that the recovered topological order respects the known stage hierarchy. The resulting order-connected graph (Def. 1) is then pruned by training, for each node, an STG-NN whose stochastic gates (Eq. 4) perform L0-relaxed variable selection; edges whose final gate collapses to zero are discarded (Algorithm 1, lines 17–20). Experiments on synthetic layered Erdős–Rényi graphs (few-stage and many-stage), the same graphs with noisy stage labels or mixed continuous/binary variables, and the expert-defined causalAssembly manufacturing simulator show that OCDM attains lower topological-order divergence and better AUROC/AUPRC/SHD/SID than PC, NOTEARS-MLP, DAG-GNN, GraN-DAG, CAM, SCORE, DAS and vanilla DiffAN.","tokens_in":17098,"tokens_out":1090,"duration_ms":8266,"significance":"Multistage manufacturing, service and information systems are ubiquitous, yet standard causal-discovery algorithms ignore the known stage order and scale poorly. By injecting stage membership into an order-based search and replacing CAM-style spline pruning with a more flexible neural gate, OCDM supplies a practical, computationally competitive pipeline that demonstrably improves recovery of layered DAGs on both controlled synthetic graphs and a realistic industrial simulator. The empirical gains (especially the large reduction in D_top and the favorable SID on causalAssembly) are concrete and reproducible; if the method’s assumptions hold, the work offers a useful tool for root-cause analysis and process optimization in multistage settings.","major_comments":[{"comment":"Sec. III-D and Algorithm 1 (lines 17–20) replace CAM’s group-LASSO spline selection with STG-NN pruning (Eq. 4). No consistency or parent-set recovery result is supplied under the nonlinear additive model of Eq. 1. The authors themselves note in Sec. V that both STG-NN and CAM “lack theoretical guarantees for recovering the true parent set” and that over-pruning of true edges remains a problem. Consequently the multi-metric gains reported in Figs. 3–5 could be driven by aggressive sparsity rather than faithful recovery of the layered DAG; a modest recovery guarantee, or at least a controlled ablation that quantifies true-positive versus false-positive edge retention as candidate-set size grows, is needed to substantiate the central claim that OCDM recovers the causal structure.","section":null},{"comment":"The leaf-identification argument (Var[∂s_i/∂x_i]=0) and the stage-wise search both rest on the additive Gaussian noise model (Eq. 1) and on perfectly known stage labels (Sec. III-A). While the noisy-label and mixed-type experiments (Figs. 6–7) are welcome, they remain limited: noise is injected only by random reassignment of a fraction of labels, and the mixed-type data still use a logistic transform of the same Gaussian-process functions. A clearer statement of the breakdown regimes—e.g., when stage information is only partially known or when non-additive interactions dominate—would strengthen the paper’s applicability claims.","section":null}],"minor_comments":[{"comment":"Fig. 1 caption refers to “Figures 2 and 3 of [36]”; the adapted figure itself is hard to read at the printed scale—consider enlarging node labels or providing a higher-resolution version.","section":null},{"comment":"The computational-complexity claim for STG-NN pruning (O(n·E·L·m²) vs. O(nk²+k³) for CAM) is stated without reporting wall-clock numbers for the low-dimensional Simulation 1; adding those numbers would make the efficiency discussion more complete.","section":null},{"comment":"Typographical inconsistencies appear in the arXiv identifier (2607.03971) and in the repeated phrase “to address theses issues” (p. 1); a careful proof-reading pass is warranted.","section":null},{"comment":"Hyper-parameter settings for the diffusion model and the STG regularization schedule are deferred to the supplement; a short summary table in the main text would improve reproducibility.","section":null}],"recommendation":"major_revision","confidential_remarks":"The core algorithmic idea—stage-constrained order search—is sound and the empirical package is stronger than many causal-discovery submissions. The missing recovery analysis for STG-NN is the only load-bearing gap; once that is addressed (even with a limited theorem or a thorough ablation), the paper should be publishable. Scope fits TKDE well."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The real contribution here is practical and clear: they force DiffAN’s leaf search to run stage-by-stage from the last stage upward, so the recovered order respects known process hierarchy, then swap CAM’s spline pruning for STG-NN gates. That combination is new, and the numbers back it up. On both the 3-stage and 20-stage synthetic layered graphs and on the expert-defined causalAssembly simulator, OCDM posts the lowest topological-order divergence and the best AUROC/AUPRC/SHD/SID package. The noisy-stage and mixed-type ablations are honest and show the method stays competitive until stage labels get badly scrambled. Runtime is also competitive once you hit 200 nodes.\n\nWhat they do well is keep the claims proportional. They never pretend this is a new identification theory; they just show that injecting the stage partition into an existing score-matching order search removes the counter-intuitive reverse edges that generic methods produce. The STG-NN pruning step is presented as an engineering upgrade that scales better than group-LASSO splines on high-dimensional candidate sets, and the wall-clock numbers support that claim.\n\nSoft spots exist but are not load-bearing. The leaf-identification argument still rests on the nonlinear additive Gaussian model (Eq. 1) and known stage membership; both are stated up front. The pruning step itself has no parent-set recovery guarantee—exactly as the authors note in the limitations—so some of the multi-metric gains could come from aggressive sparsity rather than perfect recovery. That is a real caveat, but it is shared with every other order-based method that uses CAM-style pruning, and the empirical edge over those methods remains. No code is shipped, which is annoying for a methods paper, yet the experimental design is transparent enough that a careful reader can still judge the results.\n\nThis is for people who actually run causal discovery on manufacturing lines, multi-stage service systems, or any layered process where stage labels are cheap and reliable. It is not a theory paper and will not reorganize the field, but it is a clean, usable increment. I would send it to referees; the contribution is real and the evaluation is thorough within its stated model.","headline":"Solid engineering fix for multistage causal discovery: stage-constrained DiffAN order search plus STG-NN pruning beats baselines on layered graphs and causalAssembly, with the usual additive-noise caveats.","tokens_in":17701,"tokens_out":558,"would_cite":true,"duration_ms":5620,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Knowing which stage each variable comes from lets causal discovery recover multistage process graphs more accurately than methods that ignore that structure.","keywords":["causal discovery","multistage processes","causal order search","layered DAGs","score matching","stochastic gated neural networks","pruning"],"falsifier":"Generate layered multistage data from a non-additive or non-Gaussian noise model (or deliberately scramble a large fraction of stage labels) and check whether OCDM’s topological-order divergence and structural Hamming distance still beat the best uninformed baselines; if they do not, the claimed advantage disappears.","tokens_in":17671,"feed_emoji":"🏭","tokens_out":611,"duration_ms":5789,"temperature":0.7,"pith_summary":"Multistage processes—manufacturing lines, multi-step services, layered systems—generate data whose variables already carry a known sequential hierarchy: later stages cannot cause earlier ones. Standard causal discovery algorithms ignore that hierarchy, so they often invent edges that run backwards through the process and become expensive as the number of stages grows. This paper introduces OCDM, an order-based method that first recovers a causal ordering stage by stage (starting from the last stage and working forward) and then prunes the resulting dense “order-connected” graph with stochastic-gated neural networks. On both synthetic layered graphs and a near-real manufacturing simulator, the stage-aware order plus the neural pruning recovers the true causal graph more faithfully, and usually faster, than existing constraint-based, score-based and order-based baselines. The practical payoff is a causal map that respects the known process layout and can therefore be trusted for root-cause analysis or intervention planning.","feed_headline":"Stage labels turn multistage causal discovery accurate","feed_subtitle":"Order search that respects known process stages beats methods that treat variables as unstructured","key_machinery":"Stage-wise leaf identification inside a DiffAN-style score-matching order search, producing an order-connected graph that is then sparsified by stochastic-gated neural networks (STG-NN) whose continuous Bernoulli gates perform L0-style parent selection.","core_discovery":"When every variable is known to belong to a stage of a layered process, a score-matching order search that is forced to identify leaves only inside the current stage, followed by stochastic-gated neural-network pruning of the resulting order-connected graph, recovers the underlying directed acyclic graph more accurately than methods that treat the variables as an unstructured bag.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Stage-locked order search recovers multistage DAGs better","Process stages force leaf order for accurate causal graphs","OCDM ties stage knowledge to score-matching order inference","Gated pruning after stage-aware order beats bag-of-vars methods","Incorporating stage labels into causal order search cuts errors"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The data must come from a nonlinear additive Gaussian noise model whose variables already sit in known stages of a layered DAG; if either the additive-noise form or the stage labels are badly wrong, both the leaf test and the stage-wise search break.","fun_headline_variants_meta":{"raw":{"variants":["Stage-locked order search recovers multistage DAGs better","Process stages force leaf order for accurate causal graphs","OCDM ties stage knowledge to score-matching order inference","Gated pruning after stage-aware order beats bag-of-vars methods","Incorporating stage labels into causal order search cuts errors"]},"model":"grok-4.5","effort":"low","cost_usd":0.0039,"raw_usage":{"total_tokens":1166,"prompt_tokens":774,"num_sources_used":0,"completion_tokens":85,"cost_in_usd_ticks":39000000,"prompt_tokens_details":{"text_tokens":774,"audio_tokens":0,"image_tokens":0,"cached_tokens":0},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":307,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":774,"tokens_out":85,"duration_ms":3033,"temperature":1.0,"reasoning_tokens":307,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-11T22:37:29.524353+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Generate layered multistage data from a non-additive or non-Gaussian noise model (or deliberately scramble a large fraction of stage labels) and check whether OCDM’s topological-order divergence and structural Hamming distance still beat the best uninformed baselines; if they do not, the claimed advantage disappears.","supporting_citations":[],"review_version":1}