{"id":"ad779871-d078-47b3-b2eb-248dad005ba5","arxiv_id":"2607.12535","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.5,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Recoverability-adaptive Sinkhorn transport restores sample-level survey metadata under a known prior P(M), cutting total variation to <0.001 with modest accuracy loss on CHNS, NHANES, and BRFSS.","lead":"The paper restores missing survey metadata (urban/rural, gender, diabetes) from covariates while forcing the restored labels to match known population shares. It matters because unconstrained ML can be accurate yet still warp the group proportions that health and inequality studies treat as population structure.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"The strongest claim is carried by Sinkhorn under P(M), not by recoverability-adaptive transport; the latter is not shown to improve the operating point.","rationale":"The Reader correctly isolates the weakest assumption: that R̂_i = 1 − H(p_i)/log K is a faithful enough stand-in for I(X;M) to decide when evidence should dominate the prior. The paper’s own Table 3b already shows that, under present α estimates, adaptive ε does not beat fixed η. My concern is the same soft spot, sharpened: the headline TV-restoration result is driven by the marginal constraint that any Sinkhorn readout supplies, so the distinctive recoverability story is not yet load-bearing for the empirical claim. That still warrants CONDITIONAL (not REJECT), because the population-consistency result itself is solid and useful for survey workflows; it simply needs the adaptive mechanism either demonstrated to help under stronger uncertainty signals or clearly demoted. Agreement with the Reader is therefore full on the weakest assumption and on the CONDITIONAL verdict.","tokens_in":11467,"tokens_out":593,"duration_ms":5801,"concrete_test":"On the same CHNS checkpoint used for Table 3b, recompute the accuracy–TV Pareto for (i) fixed η grid, (ii) εᵢ ∝ (1−R̂_i), and (iii) εᵢ from a better-calibrated uncertainty (temperature-scaled or ensemble entropy). If no adaptive schedule strictly dominates the best fixed η on the joint (Acc, TV) metric, the recoverability-adaptive claim does not improve the operating point and should be demoted to optional control infrastructure.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper’s strongest empirical claim (TV≈0.11 → TV<0.001 with limited accuracy loss) is almost entirely explained by the OT/Sinkhorn column constraint under known c = P(M). Classical baselines with the same Sinkhorn post-processing already reach comparable constrained accuracy and near-zero TV (Table 2b: XGB+SK 69.51% / 81.59% / 82.66% vs CDR+SK). Recoverability-adaptive ε is the distinctive mechanism advertised in the title and abstract, yet Table 3b shows sample-specific εᵢ(α) matches fixed global η on both accuracy (≈69.1%) and TV (0.0004). The stratified analysis (Table 2e) only shows that low-R̂ rows are reassigned more often under a shared prior—it does not show that modulating ε improves the accuracy–fidelity frontier relative to fixed ε. Thus the load-bearing novelty claim (recoverability-adaptive transport as the regulator that makes CDR work) is not secured by the reported experiments; the TV restoration would hold under plain Sinkhorn.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper formalizes Context Distribution Restoration (CDR): recovering incomplete survey metadata M from covariates X under a known external prior P(M)=c, so that sample-level assignments remain informative while the reconstructed margin matches c. Recoverability is defined as R(M|X)=I(X;M) and approximated by calibrated predictive entropy R̂_i=1−H(p_i)/log K; this quantity is used to set sample-specific Sinkhorn entropies ε_i inside an entropy-regularized optimal-transport plan whose column marginal is constrained to c. Context anchors (EMA prototypes) supply the cost matrix C. On CHNS urban/rural, NHANES gender, and BRFSS diabetes (up to 67k test samples, ρ=0.1 labels retained), unconstrained classifiers achieve high accuracy but large total variation from P(M) (TV≈0.11 on CHNS), while Sinkhorn under c restores TV<0.001 with modest accuracy loss. Ablations, mask-sensitivity sweeps, recoverability-stratified reassignment rates, and classical baselines with the same Sinkhorn readout are reported, together with a CHNS continuum case study.","tokens_in":11806,"tokens_out":1310,"duration_ms":11541,"significance":"If the central empirical claim holds, CDR supplies a practical restoration layer for computational social science workflows that need both individual metadata assignments and population-consistent margins—something ordinary classification, multiple imputation, and reweighting/raking do not jointly guarantee. The multi-survey evaluation (CHNS, NHANES, BRFSS), multi-seed tables, Argmax-vs-Sinkhorn TV comparisons, mask-sensitivity analysis, and classical baselines with identical Sinkhorn post-processing are genuine strengths and make the margin-restoration result reproducible and falsifiable. The problem framing itself (sample-level restoration under hard external P(M)) is a useful contribution even if the distinctive adaptive-ε mechanism is not yet shown to improve the operating point.","major_comments":[{"comment":"The title, abstract, and §1.3/§3.4 present recoverability-adaptive transport (per-sample ε_i from R̂ or α) as the core mechanism that regulates the accuracy–fidelity trade-off. Table 3b, however, shows that sample-specific ε_i(α) matches fixed global η on both Sinkhorn accuracy (≈69.1%) and TV (0.0004). Table 2b further shows that XGBoost+Sinkhorn and LogReg+Sinkhorn already achieve near-zero TV and constrained accuracy comparable to CDR. The load-bearing novelty claim therefore rests almost entirely on the OT column constraint, not on recoverability adaptation. Either demonstrate a setting (stronger miscalibration, multi-class M, uncertain c, or a clear accuracy–TV Pareto improvement) where adaptive ε strictly dominates fixed ε, or reframe the contribution so that Sinkhorn under P(M) is primary and adaptive ε is optional control infrastructure.","section":null},{"comment":"§3.2 defines R(M|X)=I(X;M) and then substitutes the operational proxy R̂_i=1−H(p_i)/log K. Table 2e shows that low-R̂ rows are reassigned more often under a shared prior, which is consistent with the design but does not validate that the proxy tracks mutual information or improves the frontier relative to fixed ε. Because the adaptive mechanism is advertised as central, the manuscript needs either (i) a calibration/MI diagnostic linking R̂ to I(X;M) or held-out recoverability, or (ii) an explicit demotion of the proxy to an engineering schedule whose empirical benefit is not yet established (as Table 3b already suggests).","section":null},{"comment":"Eq. (2) writes a joint objective min_T ⟨T,C⟩+λ1 D(T1∥P(M))+λ2 L_recover, yet §3.4–3.5 implement standard entropy-regularized Sinkhorn at inference with column constraint c and optional per-row ε_i; training is masked CE plus MMD/α auxiliaries. The relationship between the stated objective and the implemented pipeline is therefore unclear: are λ1,λ2 optimized, or is D enforced hard by Sinkhorn? Clarifying this (and reporting whether end-to-end differentiation through Sinkhorn is used) is needed for the method section to support the claims.","section":null}],"minor_comments":[{"comment":"Abstract and §4 claim “minimal accuracy loss” while Table 2 shows CDR below unconstrained XGBoost by 2–3 pp and, on NHANES, below XGB+SK; quantify the trade-off more carefully in the abstract.","section":null},{"comment":"Table 2 vs Table 2b: CDR Sinkhorn numbers differ slightly across tables (e.g., CHNS 69.28% vs 68.95%); reconcile seeds/early-stopping protocols so the main comparison is unambiguous.","section":null},{"comment":"Figure 1 is described but the architecture diagram is not fully self-contained in the text; ensure encoder, anchors, α head, and Sinkhorn block are labeled consistently with Eqs. (5)–(7).","section":null},{"comment":"§4.6 continuum bands and APA (Supp. Table S1) are correctly demoted as non-core, yet still occupy substantial space; a shorter pointer would reduce distraction from the main claim.","section":null},{"comment":"Notation: R(M|X)=I(X;M) is written as a function of the pair but used as a per-sample scalar; a short remark that the operational proxy is sample-wise would avoid confusion.","section":null},{"comment":"Funding / author contributions / acknowledgements are still placeholders; complete before camera-ready.","section":null}],"recommendation":"major_revision","confidential_remarks":"The empirical core (OT restores TV under known P(M) on three large surveys) is solid and publishable after reframing. The risk is overselling recoverability-adaptive transport relative to plain Sinkhorn; if the authors cannot show a clear win for adaptive ε, the paper remains a useful applied OT-for-surveys contribution but should not lead with that mechanism. Scope fit for a computational social science / data-science venue is good; less so for a pure ML theory venue."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The thing to know: this paper’s real result is that entropy-regularized OT under an external prior c = P(M) drives TV from ~0.11 down to <0.001 on three national surveys with only modest accuracy loss. That claim is cleanly shown. The recoverability-adaptive transport advertised in the title and abstract is not what carries the numbers.\n\nWhat is new is the packaging for computational social science: incomplete stratification fields (urban/rural, gender, diabetes) under known census/design margins, sparse labels, and a restoration layer that must stay informative at the case level. They run multi-seed tables on CHNS, NHANES, and BRFSS (up to 67k test), put XGBoost and LogReg through the same Sinkhorn readout, do mask-sensitivity sweeps, and stratify reassignment by predictive-entropy recoverability. The Argmax-vs-Sinkhorn TV comparisons and Table 2b are the useful core. Classical models + Sinkhorn already get near-zero TV and comparable constrained accuracy; the OT column constraint is doing the work they claim for CDR.\n\nSoft spots, in proportion. Table 3b shows sample-specific εᵢ(α) matches fixed global η on both accuracy (~69.1%) and TV (0.0004). So the distinctive mechanism does not improve the frontier under their own uncertainty estimates. The stratified table only shows low-R̂ rows get reassigned more often under a shared prior—not that modulating ε helps. Continuum bands and APA are correctly demoted. Code is request-only. Recoverability as 1 − H(p)/log K is a standard uncertainty proxy, not a verified stand-in for I(X;M). None of that sinks the TV-restoration result; it just means the paper oversells the adaptive story relative to the evidence.\n\nMath and citations are ordinary and fine: Cuturi Sinkhorn, prototypes as costs, survey missing-data and raking literature. No circularity—the prior is external. This is for people who restore incomplete survey metadata before stratified analysis. It deserves a serious referee; the empirical claim is real and the framing is useful even if the novelty claim needs to be narrowed to “OT under P(M) as a restoration layer, with recoverability as optional control infrastructure.” I would engage, cite the TV result if I work on survey pipelines, and push them to either show adaptive ε wins or demote it.","headline":"Solid applied demo that Sinkhorn under known P(M) restores survey metadata margins; the recoverability-adaptive piece in the title is not yet shown to move the operating point.","tokens_in":12370,"tokens_out":589,"would_cite":true,"duration_ms":5238,"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":"Survey metadata can be restored case by case without warping known population shares.","keywords":["Context Distribution Restoration","computational social science","metadata restoration","recoverability","optimal transport","social surveys","Sinkhorn","population margins"],"falsifier":"On a held-out survey with known ground-truth labels and a trusted prior, compare sample-specific entropy schedules against a fixed global entropy schedule: if adaptive softness never improves the accuracy-fidelity frontier beyond fixed Sinkhorn under the same prior, the recoverability mechanism is not doing load-bearing work.","tokens_in":12329,"feed_emoji":"📊","tokens_out":899,"duration_ms":7557,"temperature":0.7,"pith_summary":"Large health and social surveys often lose key labels such as urban/rural status, gender, or disease flags to privacy rules, nonresponse, or linkage errors. Analysts still know the population shares of those labels from census or design margins. This paper treats the job as Context Distribution Restoration: assign a label to every incomplete case from its covariates, but force the reconstructed group shares to match the known prior. The hard part is that some cases are easy to recover from their covariates and others are ambiguous. The authors define that recoverability as mutual information and approximate it with calibrated predictive uncertainty, then use a recoverability-adaptive optimal-transport layer so easy cases follow the model and hard cases lean on the population prior. On three national surveys with up to 67k test cases, unconstrained classifiers hit high accuracy yet distort margins (total variation around 0.11), while the constrained method brings total variation below 0.001 with only modest accuracy loss. A China Health and Nutrition Survey case study also surfaces a continuum between urban and rural prototypes.","feed_headline":"Restore missing survey labels without warping group shares","feed_subtitle":"On three national surveys, transport under known priors cuts margin error below 0.001 with small accuracy cost.","key_machinery":"Recoverability-adaptive transport: Sinkhorn optimal transport under the known column prior P(M), with sample-specific entropic softness set from an operational recoverability score (one minus normalized predictive entropy) so high-recoverability rows follow model costs and low-recoverability rows defer toward the prior.","core_discovery":"The authors formalize incomplete survey metadata restoration under a known prior P(M) as Context Distribution Restoration: sample-level assignments from covariates X that remain informative when signals exist yet match the known population margin. Across CHNS, NHANES, and BRFSS, unconstrained high-accuracy classifiers can still produce large margin errors (TV approximately 0.11 on CHNS), while an optimal-transport readout under the prior restores TV below 0.001 with limited accuracy cost; the marginal constraint is the main restorer of population fidelity, and recoverability only modulates how hard individual predictions are allowed to pull against that constraint.","pith_inferences":["The same margin-constrained layer could restore multi-category administrative codes (region, education, occupation) once multi-class priors are reliable.","If official margins themselves carry error, replacing hard column constraints with soft probabilistic marginals would be the natural next stress test.","Downstream re-analysis of published disparity indices under constrained versus unconstrained restored labels would show whether the fidelity gain changes substantive conclusions."],"forward_implications":["Incomplete stratification fields can be filled for subgroup and inequality analyses without inventing distorted group sizes.","Any strong classifier can be turned into a population-consistent restorer by adding a Sinkhorn readout under known margins.","When labels are sparse, the prior correctly dominates ambiguous cases; denser labels shrink the accuracy trade-off but still need the margin constraint.","Prototype costs can expose continuum structure (for example peri-urban bands) without treating that structure as an accuracy objective."],"fun_headline_variants":["Keep survey margins exact while recovering missing labels","Transport restores labels matching known population priors","Recover missing fields without distorting group shares","Adaptive OT cuts margin TVD below 0.001 in surveys","Restore incomplete survey metadata under exact priors"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The method assumes that a simple predictive-entropy score is a good enough stand-in for true mutual information to decide, case by case, whether the model or the population prior should win.","fun_headline_variants_meta":{"raw":{"variants":["Keep survey margins exact while recovering missing labels","Transport restores labels matching known population priors","Recover missing fields without distorting group shares","Adaptive OT cuts margin TVD below 0.001 in surveys","Restore incomplete survey metadata under exact priors"]},"model":"grok-4.5","effort":"low","cost_usd":0.009442,"raw_usage":{"total_tokens":2195,"prompt_tokens":836,"num_sources_used":0,"completion_tokens":71,"cost_in_usd_ticks":94420000,"prompt_tokens_details":{"text_tokens":836,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1288,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":836,"tokens_out":71,"duration_ms":9737,"temperature":1.0,"reasoning_tokens":1288,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-15T05:19:19.456771+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On a held-out survey with known ground-truth labels and a trusted prior, compare sample-specific entropy schedules against a fixed global entropy schedule: if adaptive softness never improves the accuracy-fidelity frontier beyond fixed Sinkhorn under the same prior, the recoverability mechanism is not doing load-bearing work.","supporting_citations":[],"review_version":1}