{"id":"8dd71454-c130-4081-a205-78a12b2da7bd","arxiv_id":"2607.27867","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Orca extends structural causal models to continuous time with neural operators, enabling resolution-invariant dose-response and counterfactual trajectories on irregularly sampled cyclic systems.","lead":"The paper introduces Orca, a framework that treats each variable in a causal model as a continuous function of time and learns the causal mechanisms between variables as neural networks that map functions to functions. It supports irregular sampling, feedback loops, and answers what-if questions by re-simulating the learned system under changed conditions.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Identifiability gap acknowledged but load-bearing: no guarantee connects predictive fit to causal correctness; non-additive coupling is unproven. CONDITIONAL verdict remains appropriate.","rationale":"The reader's weakest assumption — that the operator parameterization must mirror the true mechanism — is exactly the load-bearing concern, and the paper itself states it in the Limitations: 'even with the correct graph, predictive fit does not certify a causal decomposition: the parameterization must mirror the true mechanism, or causal estimates can be badly biased while the training loss barely moves.' This is not a flaw in the paper's internal logic; it is a scoping statement. Orca is presented as a framework with synthetic feasibility demonstrations, not as a method with unconditional guarantees. The experiments support the feasibility claims: the fixed-kernel variant tracks ground truth, the universal operator tracks it in one task, and the non-additive counterfactual experiment works under a known parametric mechanism. However, the general non-additive algorithm is only sketched and unproven, and the universal operator's performance under misspecification is untested. These are real limitations, but they are openly acknowledged and do not invalidate the framework's contribution. A CONDITIONAL verdict — accept as a promising framework while treating the identifiability and coupling issues as open problems — is the right calibration. Therefore my stress-test does not change the reader's verdict.","tokens_in":13209,"tokens_out":7161,"duration_ms":91337,"concrete_test":"Add a misspecification arm to Experiment 1: generate the true health mechanism with an interaction term that the sum-of-embeddings deep-set operator cannot represent, e.g., H_i(t) = H_i(0) + ∫_0^t K(t−s)E_i(s)ds + c E_i(t)·A_i(t) (or a product of two parent integrals), fit Orca universal on observational data, and compare its dose–response curve E[H(12)|do(E=e)] to ground truth. If the training loss is low but the dose–response is visibly biased, the identifiability gap is confirmed to be practically relevant; if not, the universal operator is more robust than the paper claims.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing concern, also flagged by the paper's own Limitations section, is that predictive fit does not certify a correct causal decomposition. The central claim — that Orca supports causal effect estimation and counterfactual reasoning — holds only if the operator parameterization mirrors the true mechanism closely enough. No identifiability theorem is provided: the universal approximation results (DeepSets, MIONet) concern representation, not whether the empirical-risk minimizer yields correct interventional or counterfactual distributions under distribution shift. The non-additive coupling procedure in 'Learning Distributions over Functions' is explicitly a heuristic: it uses optimal transport on rank trajectories, but no consistency or identifiability result is given, and Experiment 2 uses a parametric ODE with known Gompertz form, not the general deep-set operator, so the framework's general non-additive capability is not actually demonstrated. This is not an internal inconsistency — the paper is candid — but it is a genuine correctness risk. The experiments only test favorable parameterizations, leaving the scope of the central claim uncertain.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Orca, a framework for causal reasoning over continuous-time processes observed at irregular time points. Each node of a causal graph is a function of time, and each mechanism is a learned neural operator mapping parent histories to node values, with latent noise treated as a function. The authors formalize a cyclic continuous-time generative model, a deep-set-style operator architecture with quadrature weighting for resolution consistency, and interventional and counterfactual inference via simulation. They demonstrate the approach on two synthetic experiments: a glucose–insulin model with a confounder and cycle, testing dose–response convergence under varying sampling density, and a tumor-growth model with non-additive latent noise, testing individual counterfactuals. The paper is candid about limitations, including the assumption of a known graph and the lack of a guarantee that predictive fit certifies a correct causal decomposition.","tokens_in":13510,"tokens_out":2774,"duration_ms":32163,"significance":"If the central claims hold, Orca would be a meaningful step toward causal inference for continuous-time, irregularly sampled, cyclic systems, connecting neural operator learning with structural causal modeling. The paper's strengths include a clean formal model (Equations 4–5), a concrete architecture with a clear resolution-consistency motivation, fully reproducible synthetic experiments with code, and an unusually candid limitations section. The universal-approximation arguments are used appropriately as motivation rather than as identifiability theorems. However, the significance is tempered by the fact that causal correctness is established only under favorable parameterizations, and the general non-additive method is explicitly a sketch. The contribution is therefore best read as a framework proposal with proof-of-concept evidence, not as a validated general method for causal effect estimation or counterfactual reasoning.","major_comments":[{"comment":"The load-bearing identifiability gap is acknowledged but not resolved. The paper states in Limitations that 'even with the correct graph, predictive fit does not certify a causal decomposition: the parameterization must mirror the true mechanism, or causal estimates can be badly biased while the training loss barely moves.' This directly affects the central claim that Orca 'supports causal effect estimation, counterfactual reasoning, and the learning of mechanistic functions.' Equation (5) trains only on observational MSE; there is no theorem or formal condition under which minimizing this loss yields correct interventional or counterfactual distributions under distribution shift. The universal approximation results cited concern representation capacity, not whether the empirical-risk minimizer recovers the correct mechanism. I would like the authors to either prove such a condition for","section":"Assumptions and Limitations"},{"comment":"The non-additive noise coupling procedure is explicitly a heuristic, and Experiment 2 does not demonstrate the general deep-set operator in this setting. The algorithm in 'Learning Distributions over Functions' uses optimal transport on rank trajectories, but no consistency, identifiability, or even convergence result is given for this coupling. In Appendix B, the tumor model is fit as a parametric ODE with the Gompertz form assumed known in advance (Equation 6) and only four constants learned, not as the universal deep-set operator described in Section 'Our Method'. Thus the claimed general non-additive capability — the main innovation over additive residuals — is not actually exercised by the experiment. The authors should either implement and test the deep-set operator on a non-additive mechanism, or clearly state that the non-additive method is only demonstrated for a parametric, nea","section":"Learning Distributions over Functions / Experiment 2"},{"comment":"The resolution-invariance claim is established only informally. In 'Semantics', the continuous-time model is defined as the weak limit of grid-based distributions as the grid is refined, with the caveat that 'avoiding pathological cases remains the modeler's responsibility.' In 'Resolution Consistency', convergence of the quadrature-weighted sum to an integral under mesh refinement is invoked for predictions. However, the paper does not prove that the interventional distributions computed by simulating the learned operator converge as the grid refines, nor that the quadrature error stays controlled when the operator is composed with downstream causal queries. Since the central selling point is that Orca decouples causal conclusions from the sampling grid, this is a load-bearing gap. A formal statement with sufficient regularity conditions on the mechanisms and noise processes, at least f","section":"Semantics / Resolution Consistency"}],"minor_comments":[{"comment":"In the illustrative example, Brownian motion is written as a term added to the trajectory, but in the formal model (Equation 4) the noise enters as an increment at each time point. The relationship between these two uses of Brownian motion could be clarified, particularly whether the equations are interpreted in an Itô or pathwise sense.","section":"Notation, Equations (1)-(2)"},{"comment":"The caption describes the neural operator baseline as 'no causal graph (no age)', but the method is better characterized as 'causally blind' rather than 'no graph', since it still uses an operator. Minor wording issue.","section":"Figure 3 caption"},{"comment":"The notation S_i^v(t) is defined informally as 'the part of X that the mechanism is allowed to see.' It would be helpful to state explicitly that X includes subject index, node index, time, and value, and that the history set is finite because measurements are finite.","section":"Section 'Data and Objective'"},{"comment":"The statement that the learned sensitivity slope is 'close to the true noise scale (w≈0.5 against σ=0.55)' is presented without error bars or a sensitivity analysis. Given that this parameter directly controls counterfactual divergence, reporting its variability across random seeds would strengthen the result.","section":"Appendix B"}],"recommendation":"major_revision","confidential_remarks":"The reader's conditional verdict is appropriate. The paper is honest and well-structured, but the central claim of general causal effect estimation and counterfactual reasoning is broader than what the experiments and theory support. The main issues — the missing identifiability guarantee, the heuristic non-additive coupling, and the parametric special case used in Experiment 2 — are all fixable within the manuscript's scope, either by adding theory or by carefully narrowing the claims. I would not reject, but I would require a substantive revision before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nYou should know this paper first for what it is: a framework proposal, not a solved causal inference method. Its real contribution is the combination — neural operators as mechanism class for continuous-time SCMs with cycles, temporal priority, multiple parent functions, and noise abducted as a function for counterfactuals. That combination is genuinely new relative to the dynamic-SCM literature (which gives semantics but no learning recipe) and neural-operator literature (which gives learning but no intervention semantics). The formalization is clean: the forward sweep unrolls cycles into a DAG, additive noise makes training a regression, and the quadrature weighting gives resolution-invariant training. The synthetic dose-response experiments show the expected convergence and the baselines fail in the expected ways. I take the paper at its word that this is feasibility, not application.\n\nThe paper also deserves real credit for its limitations section. It says plainly that predictive fit does not certify a causal decomposition and that the parameterization must mirror the true mechanism. That is the load-bearing assumption, and it is not proven. The universal approximation results (DeepSets, MIONet) are about representation, not about whether the empirical risk minimizer recovers correct interventional or counterfactual distributions under distribution shift. So the reader's weakest-assumption concern is right, and the paper's own text confirms it.\n\nThe second soft spot is the non-additive counterfactual procedure. The optimal-transport coupling of residuals to noise trajectories is explicitly a heuristic, with no consistency or identifiability result. Experiment 2, moreover, uses a parametric ODE with known Gompertz form — not the general deep-set operator — so the framework's general non-additive capability is not actually demonstrated. That is a real gap, though the paper doesn't oversell it.\n\nMinor concerns: all validation is synthetic, the data model assumes noise-free synchronous measurements, and the graph is given. These are stated. Nothing here is hidden.\n\nVerdict: this is a serious, honest framework paper. I'd send it to peer review. The right outcome is acceptance with the understanding that the identifiability question and the non-additive coupling are open problems, not solved. I'd cite it for the framework, not for any empirical claim.\n\nBring it to reading group? Yes — it will generate good discussion about what a causal operator model would need for real EHR use.","headline":"A candid, well-scoped framework paper that combines neural operators with causal semantics; the main risk is the acknowledged but unproven identifiability assumption, so treat as conditional.","tokens_in":13923,"tokens_out":1851,"would_cite":true,"duration_ms":18078,"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":"Orca treats continuous-time causal reasoning as learning maps between function spaces, enabling dose-response and counterfactual queries on irregularly sampled trajectories.","keywords":["causal reasoning","neural operators","continuous time","structural causal models","counterfactuals","dose-response","irregular time series","latent noise"],"falsifier":"Simulate a continuous-time system with a known graph and a true mechanism deliberately outside the deep-set operator class (for example, with strong non-smooth interactions or noise that enters non-monotonically), fit Orca on dense irregular samples, and check whether the estimated dose-response curve converges to ground truth under grid refinement while the training loss stays near zero. If the two diverge—training loss falling but causal error not—the central claim that predictive fit supports causal inference in this setting is falsified.","tokens_in":13135,"feed_emoji":"🕒","tokens_out":5572,"duration_ms":44602,"temperature":0.7,"pith_summary":"This paper argues that neural operator learning provides a natural foundation for causal reasoning about continuous-time systems observed at irregular time points. It introduces Orca, a framework in which each node of a causal graph is a function of time and each mechanism is a learned map between function spaces, with temporal priority built in and cycles unrolled along time. On synthetic examples, Orca recovers interventional dose-response curves and individual counterfactual trajectories, while baselines that ignore either the causal graph or time remain biased. The central claim is that the operator formulation gives causal queries a resolution-invariant semantics: conclusions no longer depend on the sampling grid.","feed_headline":"Causal reasoning goes continuous with neural operators","feed_subtitle":"Each variable becomes a function of time, so dose-response and counterfactuals no longer depend on the sampling grid.","key_machinery":"The load-bearing object is the forward operator F_v, which computes node v's value at time t from the past histories of its parents and the current noise, implementing the temporal priority principle. Orca instantiates it with a deep-set operator: a shared embedding network maps each past parent measurement to a vector, the vectors are summed with quadrature weights (the time span each measurement stands for), and a decoder outputs the node value at the query time. The quadrature weighting turns the sum into an integral estimate, making predictions resolution-consistent and letting a model trained on one grid evaluate on another. For non-additive noise, an optimal-transport coupling between","core_discovery":"Orca's central claim is that neural operators are the natural mechanism class for continuous-time causal models. Each node's value is a function of time, and each mechanism is a learned forward operator that maps the past trajectories of its parent nodes, plus a latent noise function, to the node's value at a query time. Because parents are read strictly from the past, cyclic dependencies become acyclic when unrolled in time. Under additive noise the latent noise is recovered as a residual; under non-additive noise the paper couples observed subjects to sampled noise trajectories by rank-based optimal transport and reuses the coupled noise for counterfactuals. The paper demonstrates on synth","pith_inferences":["If the framework scales beyond synthetic examples, it suggests a general recipe for any domain—pharmacology, climate, epidemiology—where data arrive as irregularly sampled trajectories with a known causal graph: estimate mechanisms as operators, then answer interventions by simulating the intervened graph.","The paper's own limitation statement implies a concrete diagnostic: compare predictive loss against causal validation error on a held-out interventional distribution; if training loss drops while causal error does not, the parameterization is misspecified.","The non-additive noise coupling is only sketched; a natural test is to apply it to generative models with non-monotone or non-persistent noise and see whether counterfactual accuracy degrades as the true noise becomes less rank-like.","One could extend the construction from functions of time to space-time fields, as the paper notes, which would make the same operator semantics applicable to spatio-temporal causal queries."],"forward_implications":["Causal conclusions from irregularly sampled longitudinal data need not be tied to a fixed time grid; a model trained at one resolution can answer queries at a finer or coarser resolution.","Feedback loops can be accommodated without instantaneous circularity by reading parents only from the past, so cyclic graphs become acyclic along time.","Latent noise functions, once inferred, support subject-level counterfactual trajectories rather than only average treatment effects.","In the fixed-kernel variant, the learned mechanism is an interpretable function of time lag, directly readable as the effect of a cause from τ time units ago.","As measurements become denser, interventional estimates converge to the continuous-time ground truth, while grid-tied or causally blind baselines stay biased at every density."],"fun_headline_variants":["Neural operators bring causality to continuous time","Continuous-time counterfactuals with neural operators","Causal AI that respects the arrow of time","From static graphs to dynamic causes","Orca: neural operators for causal time"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"Even when the directed causal graph is correct, the learned operator's parameterization must mirror the true mechanism closely enough, because a good predictive fit does not by itself certify a correct causal or counterfactual decomposition.","fun_headline_variants_meta":{"raw":{"variants":["Neural operators bring causality to continuous time","Continuous-time counterfactuals with neural operators","Causal AI that respects the arrow of time","From static graphs to dynamic causes","Orca: neural operators for causal time"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000488,"raw_usage":{"total_tokens":2214,"prompt_tokens":688,"completion_tokens":1526,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":432,"completion_tokens_details":{"reasoning_tokens":1461}},"tokens_in":432,"tokens_out":1526,"duration_ms":10747,"temperature":1.0,"reasoning_tokens":1461,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T00:00:27.331361+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate a continuous-time system with a known graph and a true mechanism deliberately outside the deep-set operator class (for example, with strong non-smooth interactions or noise that enters non-monotonically), fit Orca on dense irregular samples, and check whether the estimated dose-response curve converges to ground truth under grid refinement while the training loss stays near zero. If the two diverge—training loss falling but causal error not—the central claim that predictive fit supports causal inference in this setting is falsified.","supporting_citations":[],"review_version":1}