{"id":"2c0d86b9-1ab2-47f7-8c53-2136c36161c0","arxiv_id":"2603.14169","paper_version":2,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":6.5,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"Topological causal estimands from persistent homology identify treatment-induced shape changes under a persistent-homology ignorability condition, even when mean-based ATE/CATE are zero.","lead":"This paper proposes causal effects based on persistent homology so treatments that reshape outcome distributions without changing means can still be detected. Smart generalists may care because many real interventions change shape, multimodality, or support of outcomes rather than averages alone.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Abstract-only review cannot verify the identification proofs or the non-commutativity claim that underpins the conditional-vs-marginal distinction; those are the load-bearing steps for the central claim.","rationale":"The Reader’s weakest-assumption note correctly identifies PH-ignorability as scientifically fragile and uncheckable from the abstract, which is why the verdict is already CONDITIONAL with low confidence. My concern is adjacent but more foundational: even granting that the condition could be plausible in some domains, the abstract’s identification and non-commutativity claims are pure assertions. Without the proofs, the error-bound derivation, and the precise definitions, one cannot confirm that the conditional estimands are in fact identified or that the marginal/conditional distinction is forced by non-commutativity rather than by a modeling choice. Because the Reader already conditioned acceptance on full-text verification of theorems and estimators, no verdict change is warranted; the stress-test simply sharpens the single most load-bearing technical gap that full-text review must close first. Novelty and the mean-preserving synthetic idea remain interesting if the math holds.","tokens_in":2012,"tokens_out":554,"duration_ms":5784,"concrete_test":"Obtain the full manuscript and independently re-derive (or machine-check) the identification theorem for the conditional topological estimand under approximate PH-ignorability, confirming that the stated error bound follows from the given assumptions without additional hidden regularity on the filtration or the diagram metric; if the bound does not close or requires unstated conditions, the central claim weakens.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim rests on two technical pillars that an abstract cannot establish: (1) that topological CATE/ATE analogues are identifiable up to an explicit error bound under approximate PH-ignorability, and (2) that persistent homology does not commute with mixtures over covariates, so a marginal persistence-diagram effect is not identified from conditional PH-ignorability alone. Both are asserted as proved, yet without the statements of the theorems, the precise definition of the PH-ignorability condition (exact or approximate), the metric on persistence diagrams used for the error bound, or the argument showing non-commutativity, it is impossible to check whether the identification actually holds or whether the error bound is non-vacuous. The synthetic experiment is described only at the level of qualitative behavior (mean-based estimands near zero, topological effect large and recoverable after adjustment), so it supplies no independent verification of the identification theory. The Reader correctly flags PH-ignorability as a weak scientific assumption; the deeper load-bearing gap is that the mathematical scaffolding claimed to make that assumption usable is itself uninspectable.","agreement_with_reader":"partial"},"referee_report":{"model":"grok-4.5","summary":"The manuscript proposes a topological causal framework based on persistent homology to capture treatment-induced changes in the shape of outcome distributions that mean-based ATE and CATE can miss (e.g., unimodal-to-bimodal mean-preserving shifts). It formalizes a persistent-homology (PH) ignorability condition, defines topological analogues of CATE and ATE, and claims these estimands are identifiable up to an explicit error bound under approximate topological ignorability. It further asserts that a marginal persistence-diagram effect is not identified from conditional PH-ignorability alone because persistent homology does not in general commute with mixtures over covariates, and therefore centers the theory on conditional estimands while retaining the marginal effect as motivation. A synthetic experiment with mean-preserving topology change is reported to show near-zero mean effects alongside a large topological effect that remains recoverable after confounding adjustment.","tokens_in":2265,"tokens_out":862,"duration_ms":18533,"significance":"If the identification theory is correct with a non-vacuous error bound, and if PH-ignorability is scientifically usable, the work would give causal inference a principled way to detect and estimate effects on distributional topology that classical ATE/CATE miss. Explicitly treating non-commutativity of persistent homology with covariate mixtures is a useful conceptual clarification that prevents a common identification error. The synthetic mean-preserving topology-change design is a clear stress case for the literature. These strengths are conditional on inspectable proofs, a concrete metric and error bound, and a usable statement of PH-ignorability—none of which can be verified from the abstract alone.","major_comments":[{"comment":"The central claim is that topological CATE/ATE analogues are identifiable up to an explicit error bound under approximate PH-ignorability. Without theorem statements, the precise definition of approximate PH-ignorability, the metric on persistence diagrams, and the derivation of the error bound, it is impossible to verify that identification holds or that the bound is non-vacuous. This is load-bearing for the paper’s contribution.","section":"Abstract (identification claim)"},{"comment":"The claim that persistent homology does not in general commute with mixtures over covariates—so a marginal persistence-diagram effect is not identified from conditional PH-ignorability alone—is asserted as proved and is used to re-center the theory on conditional estimands. The abstract supplies neither the formal statement nor the argument; this step must be checkable in the full manuscript.","section":"Abstract (non-commutativity claim)"},{"comment":"PH-ignorability (exact or approximate) is the key identifying assumption. The abstract places it at the center of identification but does not indicate empirical diagnostics, domain justification, or how residual confounding that systematically alters persistence diagrams would be detected or covered by the error bound. For the framework to be usable, the manuscript needs a clear statement of when the assumption is plausible and how the bound behaves under realistic violations.","section":"Abstract (PH-ignorability)"}],"minor_comments":[{"comment":"Filtration/scale choices and the diagram distance or summary map used for the topological estimand are free parameters of the framework; the abstract does not indicate how they are fixed or sensitivity-analyzed in the synthetic experiment.","section":"Abstract"},{"comment":"The synthetic experiment is described only qualitatively (mean effects near zero; topological effect large and recoverable). Quantitative results, estimator definitions, and the confounding-adjustment procedure should be fully specified in the manuscript.","section":"Abstract (synthetic experiment)"}],"recommendation":"uncertain","confidential_remarks":"This review is based solely on the abstract; the full text was not available. A proper assessment of soundness requires the theorems, proofs, estimator construction, and experimental details. I recommend the editor obtain the full manuscript before a final decision. On the abstract alone the program is coherent and the non-commutativity clarification is promising, but the load-bearing mathematical claims cannot be verified."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The one thing to know: this is an abstract-only methods paper that proposes persistent-homology analogues of ATE and CATE under a PH-ignorability condition, with identification up to an explicit error bound and a clear warning that PH does not commute with covariate mixtures, so marginal diagram effects are not free from conditional ignorability. That non-commutativity caveat is the most careful move in the abstract; they keep the marginal quantity as motivation and put the conditional estimands at the center.\n\nWhat is new is the packaging: topological estimands defined from persistence of outcome laws, a named PH-ignorability assumption (exact and approximate), and the claim of an explicit error bound under the approximate version. Relative to standard mean-based and even quantile/distributional treatment effects, that is a real expansion of the toolkit for the mean-preserving shape-change failure mode they describe (unimodal to bimodal, same mean). The synthetic experiment is the right qualitative check: means stay near zero, topological effect rises and is recoverable after adjustment. From the abstract there is no sign of circular fitting; the program is definitional and identification-theoretic.\n\nSoft spots are exactly what you would expect with no full text. We cannot inspect the theorems, the precise metric on diagrams, the filtration choices, or whether the error bound is non-vacuous. PH-ignorability is a strong scientific assumption; the abstract does not give diagnostics or domain justification beyond the toy. Free parameters (filtration, diagram distance/summary) will matter for any estimator. The stress-test is right that the identification and non-commutativity claims are load-bearing and currently uncheckable; that is not a flaw in the argument so much as a hard limit of abstract-only review. The Reader’s moderate novelty/significance scores look fair; soundness is necessarily provisional.\n\nThis is for causal methodologists and applied people who already care about distributional effects and multimodality (genomics, economics of heterogeneous responses, certain imaging settings). It deserves a serious referee once the full paper is up—proofs, estimators, and code would decide it. I would not cite from the abstract alone, and I would not put it in reading group until we can see the math. Send it to peer review if the full text materializes with the claimed proofs; desk-reject only if the theorems are missing or the bound is empty.","headline":"Abstract-only methods pitch for PH-based ATE/CATE analogues; idea is clean and useful if the proofs hold, but we cannot check the load-bearing identification or non-commutativity claims yet.","tokens_in":2881,"tokens_out":589,"would_cite":false,"duration_ms":6268,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62D20","55N31","62G05"],"pacs":[],"model":"grok-4.5","headline":"Topological causal effects based on persistent homology detect mean-preserving changes in outcome shape that classical ATE and CATE miss.","keywords":["causal inference","average treatment effect","conditional average treatment effect","persistent homology","topological data analysis","ignorability","persistence diagrams","mean-preserving topology change"],"falsifier":"A synthetic or real data setting in which the outcome law changes topology under treatment while means stay fixed, yet after covariate adjustment the estimated topological CATE remains near zero (or exceeds the paper's stated error bound) while classical ATE stays near zero; that would refute recoverability under the claimed ignorability.","tokens_in":2892,"feed_emoji":"△","tokens_out":635,"duration_ms":5371,"temperature":0.7,"pith_summary":"Average treatment effects and conditional average treatment effects only track shifts in expected outcomes, so they report near zero when a treatment reshapes the geometry or topology of the outcome distribution without moving its mean. A standard example is a unimodal control law that becomes bimodal under treatment while the means stay identical. This paper builds a causal framework around persistent homology to capture exactly those shape changes. It introduces a persistent-homology ignorability condition (and controllable approximate versions), defines topological analogues of CATE and ATE, and proves that the conditional topological estimands are identifiable up to an explicit error bound under that condition. The authors also show that a pure marginal persistence-diagram effect is not identified from conditional topological ignorability alone, because persistent homology does not commute with mixtures over covariates; they therefore keep the marginal quantity as motivation while centering the theory on the conditional estimands. A synthetic experiment with mean-preserving topology change confirms that classical ATE and CATE stay near zero while the topological effect rises sharply and remains recoverable after confounding adjustment.","feed_headline":"Topology catches treatment effects that leave means untouched","feed_subtitle":"Persistent-homology causal estimands stay identifiable when classical ATE and CATE report zero under pure shape change.","key_machinery":"Persistent-homology ignorability (and its approximate form): a condition that residual confounding does not systematically alter the topology of the outcome law beyond a controllable error; it supplies the identification argument for the topological CATE and ATE.","core_discovery":"Under a persistent-homology ignorability condition (including approximate versions), topological analogues of CATE and ATE defined via persistent homology are identifiable up to an explicit error bound, and can detect mean-preserving topology changes that leave classical ATE and CATE near zero.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Topology flags causal shape shifts that leave ATE and CATE at zero","Persistent homology recovers mean-preserving treatment effects","Topological CATE stays identifiable when classical means do not move","Homology-based effects detect bimodal jumps missed by average outcomes","Persistent-homology ignorability yields identifiable shape-change estimands"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"That persistent-homology ignorability, or a controllable approximate version of it, is a scientifically plausible and checkable condition on how treatment, covariates, and the topology of the outcome law relate.","fun_headline_variants_meta":{"raw":{"variants":["Topology flags causal shape shifts that leave ATE and CATE at zero","Persistent homology recovers mean-preserving treatment effects","Topological CATE stays identifiable when classical means do not move","Homology-based effects detect bimodal jumps missed by average outcomes","Persistent-homology ignorability yields identifiable shape-change estimands"]},"model":"grok-4.5","effort":"low","cost_usd":0.004996,"raw_usage":{"total_tokens":1366,"prompt_tokens":752,"num_sources_used":0,"completion_tokens":87,"cost_in_usd_ticks":49960000,"prompt_tokens_details":{"text_tokens":752,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":527,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":752,"tokens_out":87,"duration_ms":5244,"temperature":1.0,"reasoning_tokens":527,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T21:26:08.439807+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"A synthetic or real data setting in which the outcome law changes topology under treatment while means stay fixed, yet after covariate adjustment the estimated topological CATE remains near zero (or exceeds the paper's stated error bound) while classical ATE stays near zero; that would refute recoverability under the claimed ignorability.","supporting_citations":[],"review_version":1}