{"id":"98e1c131-9031-4f96-b084-8a2f42d32408","arxiv_id":"2505.02212","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"Markovian triangular monotonic SCMs with a known causal order are completely counterfactually identifiable from the observational distribution.","lead":"The paper introduces exogenous isomorphism, an equivalence between structural causal models that guarantees they answer every counterfactual question identically. It shows that a broad class of triangular monotonic structural causal models is fully counterfactually identifiable from the observational distribution, and it tests this with neural network implementations.","discovery_kind":"unification","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theoretical core is mostly sound, but the empirical validation is circular: model checkpoints are selected using validation CTF RMSE, which requires ground-truth counterfactual labels, so the reported counterfactual consistency is not established from observational data alone.","rationale":"The reader's verdict of CONDITIONAL is appropriate. I read the theoretical core in good faith: exogenous isomorphism (Definition 3.1) plus the KR-transport argument (Corollary 5.4) gives a credible proof that triangular monotonic Markovian SCMs sharing the causal order and observational distribution are indistinguishable at the full counterfactual layer. The compressed proof step in Theorem 4.3 and the implicit pa(i)-versus-pr(i) conditioning in Corollary 5.4 are real gaps in presentation, but they are repairable: using prefix conditionals P_{V_i|V_{pr(i)}} avoids needing to know the exact parent sets, and the local Markov property makes these equal to parent conditionals. The experimental circularity, however, is a concrete and currently unmet burden: selecting checkpoints by validation CTF RMSE uses ground-truth counterfactual labels, so the experiments do not demonstrate that the method achieves counterfactual consistency from observational data alone. This is the single most load-bearing concern for the paper's full claim, including its empirical validation; it warrants the CONDITIONAL verdict until the experiments are re-run with purely observational model selection.","tokens_in":54144,"tokens_out":17916,"duration_ms":224918,"concrete_test":"Re-run the ER-DIAG-50 and ER-TRIL-50 experiments with model selection based only on observational criteria: select the checkpoint with the lowest validation OBS WD (or lowest validation negative log-likelihood), with no access to counterfactual labels. Then report test CTF RMSE for the observationally selected checkpoints and compare with the oracle-CTF-selected results in Table 1. If the observationally selected CTF RMSE remains close to the oracle-selected values (within the reported 95% CI), the method is validated without leakage.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central theoretical claim, Corollary 5.4 together with Theorem 3.2, appears to be a credible extension of prior identifiability results; the compressed step in Theorem 4.3 is fillable via prefix-based conditionals and does not seem fatal. The load-bearing concern is instead in the experimental validation (Section 8 and Appendix D.3): 'The model weights corresponding to the epoch with the lowest CTF RMSE on the validation set are saved for testing.' CTF RMSE is computed against ground-truth counterfactual outcomes (Appendix D.2). Thus the reported low test CTF RMSE is achieved by selecting, among all training epochs, the one that best matches ground-truth counterfactuals on the validation set. This leaks counterfactual information into model selection. The ablations in Table 1 are also affected: the non-ablated model is selected with oracle CTF information, while the comparison is meant to demonstrate that observational training alone leads to L3-consistency. The experiments therefore do not currently support the paper's claim that the method 'addresses the consistency problem in counterfactual reasoning' using only observational data. This does not refute the theory, but it invalidates the empirical support for it.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies full counterfactual identifiability, denoted ~L3-identifiability, within the Pearl Causal Hierarchy. It introduces a model-level equivalence relation called exogenous isomorphism (~EI) and proves that ~EI-identifiability implies ~L3-identifiability. The authors then give sufficient conditions for ~EI-identifiability in two classes: bijective SCMs (BSCMs), via the concept of counterfactual transport and optimal-transport uniqueness, and triangular monotonic SCMs (TM-SCMs), where the counterfactual transport is shown to coincide with the Knothe-Rosenblatt transport. The main theoretical corollary states that any Markovian TM-SCM is ~EI-identifiable from the causal order, the Markov assumption, and the observational distribution alone, which strengthens known counterfactual-outcome identifiability results to full L3-consistency. The paper also proposes neural TM-SCM implementations (DNME, TNME, CMSM, TVSM) and reports synthetic experiments that claim to validate the theory.","tokens_in":54380,"tokens_out":10137,"duration_ms":126022,"significance":"If the theory is correct, Corollary 5.4 together with Theorem 3.2 is a substantial advance: it unifies and generalizes several prior results (Lu et al., 2020; Nasr-Esfahany et al., 2023; Scetbon et al., 2024) by extending identifiability from counterfactual outcomes to the entire L3 layer. The formal apparatus is a strength: the definitions are measure-theoretically explicit, the appendix provides a detailed dependency graph and proof structure, and the authors make a credible case that their results subsume earlier counterfactual-equivalence notions while being weaker than full model identifiability. The KR-transport and TM-SCM arguments form a coherent chain. However, the empirical validation, which is advertised as supporting the practical claim of observational-only counterfactual consistency, is undermined by oracle-based model selection; the experiments therefore do not currently provide the advertised evidence. The theoretical contribution is valuable and likely defensible, but the experimental section needs substantive repair.","major_comments":[{"comment":"The experiments do not currently establish that observational training alone achieves L3-consistency. Appendix D.3 explicitly states: 'The model weights corresponding to the epoch with the lowest CTF RMSE on the validation set are saved for testing.' CTF RMSE (Appendix D.2) is computed against ground-truth counterfactual outcomes. Therefore the test CTF RMSE in Table 1, Table 6, and Figure 1 is obtained by oracle selection over training epochs, which leaks counterfactual information into model selection. The non-ablated models and all ablations are selected by the same oracle criterion, so the comparisons in Table 1 do not isolate the effect of the assumed structural condition on L3-consistency from the effect of selection. Please re-run the experiments with a model-selection rule based only on observational data (for example, validation OBS WD or the NLL), or explicitly report the performance at a fixed epoch without oracle selection.","section":"Appendix D.3 and Section 8"},{"comment":"The proof of the reverse direction of Theorem 4.3 is incomplete as written. It states 'By Lemma A.8 and Lemma A.11, we have Γ(1)=Γ(2)∘h', but Lemmas A.8 and A.11 are proved under the full exogenous-isomorphism assumptions, which include the mechanism isomorphism that the reverse direction is supposed to establish. What is needed is an induction over the common causal order, propagating the component-wise relation (f_i^(2)(v,·))^{-1}∘(f_i^(1)(v,·))=h_i to prefix potential responses; this gap appears fillable but should be spelled out. The statement should also clarify how the 'almost surely' quantification interacts with the quantification over all v∈Ω_V.","section":"Appendix A.3, Theorem 4.3"}],"minor_comments":[{"comment":"The text says that the w/o O and w/o T configurations 'fail to converge', but in Figure 1 the w/o O curves appear flat with a high CTF RMSE rather than divergent; please describe the behavior more precisely, e.g., 'the CTF RMSE does not decrease as the observational fit improves'.","section":"Section 8 and Figure 1"},{"comment":"The table has formatting issues: the best model entries (for example, DNME '-0.53 ±0.05') contain a stray leading dash that is likely a LaTeX minus-sign artifact. In addition, some confidence intervals are very wide (e.g., TNME w/o O, 11.24±20.98); the paper should comment on the instability behind these intervals rather than only reporting the means.","section":"Table 1"},{"comment":"In Definition 4.4 the notation K_M(·,v,v′) overloads the first argument: for a fixed pair (v,v′), the displayed object is a map on the third Ω_V factor. Using distinct symbols or explicitly distinguishing the three factors would improve readability.","section":"Definition 4.4"}],"recommendation":"major_revision","confidential_remarks":"The oracle-based checkpoint selection is a real blocker for the experimental claims, but the theoretical contribution seems defensible and the empirical issue is fixable by changing the selection criterion. If the authors repair the experiments and tighten the proof of Theorem 4.3, the paper would be publishable in my view. The manuscript is within scope for the venue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The theory in this paper is largely sound and genuinely extends prior results: Corollary 5.4 together with Theorem 3.2 gives full ~L3 identifiability for vector-valued triangular monotone SCMs from causal order, Markovianity, and observational distribution. That is a real step beyond the scalar counterfactual-outcome results in Lu et al., Nasr-Esfahany et al., and Scetbon et al. The exogenous isomorphism framing is a useful unifying device, and the proof route through TM mappings and KR transport is credible. The compressed step in Theorem 4.3's 'if' direction is fillable by induction over the causal order, as the reader notes; I do not see it as fatal.\n\nThe soft spot is the experiments, and the stress-test concern is right. Checkpoints are selected using validation CTF RMSE, which is measured against ground-truth counterfactual outcomes. That leaks counterfactual information into the reported test numbers, and it affects every ablation because the non-ablated model enjoys oracle-based selection while the comparison is meant to show observational training suffices. So the empirical section does not currently support the paper's claim that the method 'addresses the consistency problem in counterfactual reasoning' from observational data alone. This is a serious flaw in the empirical validation, but it does not damage the theoretical core. The theory may still be correct; the experiments just do not establish it.\n\nMinor quibble: the related-work framing is a bit generous in places, but the citation pattern is honest and the formal results are genuinely new. The proof of Theorem 4.3 should be expanded in revision, and the experimental protocol needs to change (e.g., select checkpoints on validation likelihood or observational fit, then report test CTF RMSE).\n\nWho is this for? Researchers working on counterfactual identifiability, neural SCMs, and normalizing flows will get real value from the theoretical parts. It deserves a serious referee. The right outcome is major revision: keep the theory, fix the experiments, and temper the empirical claims.","headline":"Solid theoretical advance on full L3 identifiability for triangular monotone SCMs, but the experiments leak ground-truth counterfactuals into model selection.","tokens_in":54952,"tokens_out":1604,"would_cite":true,"duration_ms":18961,"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":"This paper proves that Markovian triangular monotonic SCMs sharing a causal order and an observational distribution are indistinguishable at every level of the Pearl causal hierarchy, making all counterfactual questions answerable from…","keywords":["counterfactual identifiability","Pearl causal hierarchy","exogenous isomorphism","triangular monotonic SCM","counterfactual transport","KR transport","neural causal models","Markovian assumption"],"falsifier":"Generate pairs of Markovian TM-SCMs with vector-valued variables from the same causal order and identical conditional densities evaluated on a fine grid, then compute the same nested counterfactual probability, such as $P(V_1[x_1]\\in A, V_3[x_3]\\in B \\mid V_2=v)$, for each pair by exact simulation. Corollary 5.4 predicts the two values coincide up to numerical tolerance; the central claim is refuted if any such pair differs beyond tolerance. The most informative search varies the monotone mechanisms while holding all conditional densities fixed, since that is the only degree of freedom the theory says does not matter.","tokens_in":53934,"feed_emoji":"🔄","tokens_out":10304,"duration_ms":106024,"temperature":0.7,"pith_summary":"The paper targets complete counterfactual identification: guaranteeing that every structural causal model consistent with the stated assumptions returns the same answer to every Pearl-hierarchy level-3 query. To make the problem tractable it introduces exogenous isomorphism, a componentwise bijection between noise variables that preserves both mechanisms and the noise distribution, and shows that exogenously isomorphic models are indistinguishable at level 3. The central result, Corollary 5.4, is that any Markovian triangular monotonic SCM is identifiable from its causal order and observational distribution up to this isomorphism. This means two such models that match the observable world cannot disagree on any what-if question, which extends and unifies earlier identifiability theorems for bijective, monotone, and fixed-point causal models. A reader should care because it identifies a practical class of neural causal models on which every counterfactual answer, not just a single outcome, can be trusted from observational data alone.","feed_headline":"Monotone causal models fix every counterfactual answer","feed_subtitle":"With known causal order, matching observed probabilities leaves no room for different what-if answers.","key_machinery":"The load-bearing object is the triangular monotonic (TM) mapping: a coordinate-wise transformation in which component $j$ is strictly monotone in coordinate $j$ and depends only on earlier coordinates. TM maps are bijective, and they are closed under inversion, composition, and taking contiguous subcomponents, with a well-defined monotonicity signature. The crucial identity is that composing one TM map with the inverse of another of the same signature always gives a strictly increasing triangular map, and for distributions with strictly positive density such maps coincide almost surely with the KR transport. In a TM-SCM every counterfactual transport is therefore an increasing triangular map; Markovianity ties the relevant conditional distributions to the observed ones; and KR transport's uniqueness forces all models in the class to share the same counterfactual transport. That shared transport induces exogenous isomorphism, which Theorem 3.2 converts into full level-3 consistency.","core_discovery":"Within the Pearl Causal Hierarchy the counterfactual layer encodes all causal information, so two SCMs that agree on all level-3 statements are indistinguishable for every causal question. The paper's discovery is an equivalence relation, called exogenous isomorphism, that captures this kind of identifiability without forcing a unique latent representation: it requires only a componentwise bijection between the exogenous variables, with the exogenous distribution and each causal mechanism preserved through that bijection. Theorem 3.2 shows this relation implies level-3 consistency for arbitrary recursive SCMs. For bijective SCMs, Theorem 4.6 shows that fixing counterfactual transports is enough to force exogenous isomorphism, and the KR-transport version, Theorem 4.8, reduces the needed data to conditional distributions. Corollary 5.4 then gives the headline condition: a Markovian triangular monotonic SCM is identifiable, and hence completely counterfactually identifiable, from its causal order and observational distribution alone, a guarantee that goes beyond earlier counterfactual-outcome identifiability results.","pith_inferences":["The same argument should extend to any other query expressible in level 3, such as path-specific effects and probabilities of causation, because the guarantee concerns the whole theory of the model rather than a selected query; the paper states this consequence but does not single out those applications.","A testable practical extension is to estimate the causal order from data and then feed it to a neural TM-SCM; the paper's order-reversal ablation predicts that order misspecification will degrade counterfactual accuracy even when observational fit is good, so an uncertainty-aware order selection step is a natural next component.","Because exogenous isomorphism only needs the existence of a componentwise bijection, a practitioner can deliberately relabel noise variables, for instance by standardizing them, without changing any causal answer; the normalizing-flow choice in the neural models exploits precisely this freedom.","The boundary of the guarantee can be probed by replacing triangular monotone mechanisms with general bijective autoregressive mechanisms: monotonicity is what makes the counterfactual transport agree with KR transport, so outside the TM class two models with the same observational distribution can be expected to diverge on counterfactual queries."],"forward_implications":["Two Markovian TM-SCMs with the same causal order and the same observational distribution must give identical answers to every level-3 query, including nested counterfactual conjunctions, not just single counterfactual outcomes.","Earlier identifiability theorems for monotone state transitions, bijective causal mechanisms, and fixed-point causal generative models become special cases of Corollary 5.4, with endogenous variables no longer required to be scalar.","For a neural TM-SCM trained by maximum likelihood, convergence to the observed distribution is enough to guarantee that its counterfactual predictions match those of the true model, as long as the four structural assumptions hold.","Counterfactual systems do not need to recover the true noise variables; any componentwise bijective encoding that preserves mechanisms and exogenous distribution is sufficient for complete counterfactual identifiability."],"supporting_citations":[{"why":"Defines the Pearl Causal Hierarchy and its L1, L2, and L3 layers, the framework in which the paper's target notion of complete counterfactual identifiability is stated.","marker":"Bareinboim et al., 2022"},{"why":"Supplies the recursion, solvability, and solution mapping Gamma that the isomorphism and counterfactual-transport arguments build on.","marker":"Bongers et al., 2016"},{"why":"Gives existence and almost-sure uniqueness of the KR transport, which is the fixed transport used in Theorem 4.8.","marker":"Santambrogio 2015"},{"why":"Establishes that strictly increasing triangular maps coincide almost surely with the KR transport, the property used in Lemma 5.1 and Corollary 5.4.","marker":"Jaini et al. 2019"},{"why":"Earlier monotone state-transition identifiability result that the paper subsumes as a special case of Corollary 5.4.","marker":"Lu et al., 2020"},{"why":"Supplies BGM equivalence and the bijective-mechanism counterfactual identifiability theorem that Theorem 3.2 and Corollary 5.4 extend.","marker":"Nasr-Esfahany et al., 2023"},{"why":"Fixed-point causal generative model whose counterfactual identifiability result Corollary 5.4 generalizes.","marker":"Scetbon et al., 2024"}],"fun_headline_variants":["Exogenous isomorphism proves full counterfactual identifiability","One bijection between latents settles every causal question","Triangular monotonic SCMs are completely counterfactually identifiable","Observational data alone fix all counterfactual statements"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The guarantee rests on knowing the true causal order and the exact coordinate alignment of each variable, and on the true model being Markovian with every mechanism strictly monotone in its own noise; if any of these is misspecified, the transport recovered from the observational distribution is not the true counterfactual transport and counterfactual answers can be wrong.","fun_headline_variants_meta":{"raw":{"variants":["Exogenous isomorphism proves full counterfactual identifiability","One bijection between latents settles every causal question","Triangular monotonic SCMs are completely counterfactually identifiable","Observational data alone fix all counterfactual statements"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000403,"raw_usage":{"total_tokens":2100,"prompt_tokens":948,"completion_tokens":1152,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":564,"completion_tokens_details":{"reasoning_tokens":1097}},"tokens_in":564,"tokens_out":1152,"duration_ms":8736,"temperature":1.0,"reasoning_tokens":1097,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T01:00:01.084364+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate pairs of Markovian TM-SCMs with vector-valued variables from the same causal order and identical conditional densities evaluated on a fine grid, then compute the same nested counterfactual probability, such as $P(V_1[x_1]\\in A, V_3[x_3]\\in B \\mid V_2=v)$, for each pair by exact simulation. Corollary 5.4 predicts the two values coincide up to numerical tolerance; the central claim is refuted if any such pair differs beyond tolerance. The most informative search varies the monotone mechanisms while holding all conditional densities fixed, since that is the only degree of freedom the theory says does not matter.","supporting_citations":[{"cited_title":"D., Ibeling, D., and Icard, T","cited_arxiv_id":null,"evidence_quote":"Defines the Pearl Causal Hierarchy and its L1, L2, and L3 layers, the framework in which the paper's target notion of complete counterfactual identifiability is stated."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the recursion, solvability, and solution mapping Gamma that the isomorphism and counterfactual-transport arguments build on."},{"cited_title":"One-dimensional issues, pp.\\ 59--85","cited_arxiv_id":null,"evidence_quote":"Gives existence and almost-sure uniqueness of the KR transport, which is the fixed transport used in Theorem 4.8."},{"cited_title":"A., and Yu, Y","cited_arxiv_id":null,"evidence_quote":"Establishes that strictly increasing triangular maps coincide almost surely with the KR transport, the property used in Lemma 5.1 and Corollary 5.4."},{"cited_title":"A fixed-point approach for causal generative modeling","cited_arxiv_id":null,"evidence_quote":"Fixed-point causal generative model whose counterfactual identifiability result Corollary 5.4 generalizes."}],"review_version":1}