{"id":"945f6f2c-dca1-4647-a399-3bd3f4d3e8e2","arxiv_id":"2608.08424","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":5,"one_line_summary":"Rank-based conformal scores make changepoint localization sets finite-sample valid for any frozen weights and exactly invariant to monotone data transforms, transferring certified set lengths across the entire monotone orbit.","lead":"This paper introduces a family of changepoint localization scores that use only within-segment ranks, so the resulting confidence sets are valid for any frozen score and identical under any monotone transformation of the data. A generalist should read it because it turns distribution-free coverage into distribution-free efficiency, a guarantee no plug-in score offers.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified","rationale":"The reader's strongest claim, Theorem 2, is mathematically sound and fully proven. The invariance argument is airtight: every ARC score is a function of within-segment ranks, strictly increasing transforms preserve ranks, and with shared permutation randomness every p-value is identical. Consequently the law of the confidence set and its length is constant on the monotone orbit. The reader's weakest assumption (frozen score) is a stated condition, not a defect; the paper explicitly warns against adaptive selection. The main blemish the reader identified is the incomplete proof of Proposition 1. I agree this is a genuine incompleteness, but it does not bear on the central validity or invariance claims, so it does not constitute a load-bearing objection to the paper's core contribution. The verdict should remain as the reader set it: the paper is acceptable in substance but should discharge or downgrade Proposition 1 before all theoretical claims are considered complete.","tokens_in":15041,"tokens_out":19367,"duration_ms":205007,"concrete_test":"To settle whether the Proposition 1 gap is benign, independently complete or refute the two omitted steps in Appendix A.3: prove a sub-Gaussian tail bound for max_s |sum_{i<=s} C_i|/sqrt(s(m-s)) under random permutation of centred ranks, and verify the union bound over B Monte Carlo draws with B growing logarithmically; if the bound fails, the O(sqrt(log n)) permutation quantile claim is wrong and the contraction rate needs revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"After reviewing the proof of Theorem 2, I find no load-bearing flaw in the central claim. The argument depends only on each ARC score being a deterministic function of within-segment ranks, which is true for the rank-CUSUM channels, fixed combinations, and the frozen network whose input is downsampled rank paths. The coupling by shared Monte Carlo permutations gives exact equality of p-values and hence of confidence sets, so the distributional transfer across the monotone orbit follows. The only theoretical incompleteness is Proposition 1 (Appendix A.3), whose proof explicitly omits two technical steps: a sub-Gaussian tail bound for the maximum of standardized rank CUSUMs under permutation, and the union bound over B Monte Carlo draws. That proposition concerns asymptotic length contraction, is not used in Theorem 1 or Theorem 2, and the paper transparently labels it a sketch.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ARC (Augmented-Rank Conformalization), a family of scores for conformal changepoint localization that depend on the data only through within-segment ranks: rank-CUSUM location and scale channels, fixed combinations, and a lightweight neural network score trained on synthetic data and frozen before deployment. The authors prove finite-sample coverage for any frozen score (Theorem 1), prove that the entire confidence set is almost surely invariant under strictly increasing marginal transformations (Theorem 2), and thereby obtain transfer of the set-length distribution across the monotone orbit of a distribution pair. They also state an asymptotic length-contraction proposition for the location channel (Proposition 1) with a proof sketch, and support the claims with simulations, a well-log case study, and explicit limitations covering serial dependence, trend alternatives, and unconditional coverage.","tokens_in":15213,"tokens_out":12807,"duration_ms":133691,"significance":"If the results hold, ARC fills a genuine gap: conformal wrappers make coverage universal but leave set length unprotected under heavy tails and distribution shift. The rank-based construction is simple, and the transfer theorem is a clean, falsifiable statement of distribution-robust efficiency. The paper is unusually transparent: the missing steps in Proposition 1 are acknowledged, serial dependence is tested and shown to break exactness, and conditional-on-detection coverage is separated from marginal coverage. The code and result files accompany the paper, and the simulations appear reproducible. Theorems 1 and 2 are standard but correctly executed; the main contribution is conceptual and should be of interest to the conformal inference and changepoint communities.","major_comments":[{"comment":"Proposition 1 is stated as a formal result, but the proof sketch explicitly omits two substantive steps: a sub-Gaussian tail bound for the maximum of standardized rank CUSUMs under permutation, and a union bound over the B Monte Carlo draws. In addition, the proposition statement does not specify that B must grow with n, although the proof sketch requires B to grow at least logarithmically in n. As written, the proposition is not proved. Because Theorems 1 and 2 do not depend on it, this does not undermine the central claims, but the authors should either supply a complete proof with the B_n growth condition made explicit, or reclassify the statement as a conjecture or remark with a clear caveat.","section":"Section 4, Proposition 1; Appendix A.3"}],"minor_comments":[{"comment":"The text says every coverage entry lies within Monte Carlo error of the nominal 0.90, but the t3 learned-combination cell reports 0.963, which is about 3.6 standard errors above nominal under 300 replications; please either provide a multiple-testing justification for the set of cells or recheck this cell.","section":"Section 5.1, Table 2"},{"comment":"The claim of exactly coincident p-value profiles relies on sharing the Monte Carlo permutation draws between the original and transformed windows; the text is clear about this, but it may help to state explicitly that without shared randomness the equality is distributional rather than pathwise.","section":"Section 5.2, Figure 2"},{"comment":"In the definition of the scale channel, the quantities q_i, \\bar q, and v_q are used before v_q is defined; please state whether v_q is the sample variance with divisor m-1 or m, since the normalization constant affects the statistic.","section":"Section 3.2"},{"comment":"The well-log protocol centers each window on a consensus annotation, so the annotated location is candidate 24 by construction; the paper reports this, but it should be emphasized that the localization exercise is conditional on a changepoint being known to lie in the window.","section":"Section 6"}],"recommendation":"minor_revision","confidential_remarks":"The main theorems are sound and the central contribution is clear. The only substantive gap is the unproved Proposition 1, which is transparently labeled as a sketch; this is local and fixable by completing the proof or reclassifying the statement. I would be comfortable with acceptance after the requested revisions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper is worth your time. It supplies the first score family for conformal changepoint localization whose confidence-set length is invariant under strictly increasing monotone transforms of the data. That is a real gap in the literature: coverage has been universal since Dandapanthula-Ramdas and Hore-Ramdas, but efficiency has been tied to the oracle likelihood ratio. ARC makes efficiency transferable by scoring only within-segment ranks.\n\nThe central theory is sound. Theorem 1 is the standard exact permutation-p-value argument, with the quantifier that any frozen weight configuration, including random or mistrained networks, retains finite-sample coverage. That is correctly stated and correctly proved. Theorem 2 is a direct consequence of rank invariance, and the paper is honest about the coupling caveat: shared Monte Carlo permutations make the confidence sets exactly equal, and it is the law of set length that transfers across the monotone orbit. The stress-test note checked this carefully and found no load-bearing flaw; I agree. The efficiency comparisons, the sabotage experiments, the well-log analysis, and the honest reporting of serial dependence all back up the claims. The code and result files are a real asset; the paper is reproducible in a way many conformal papers are not.\n\nThe soft spots are real but localized. Proposition 1, the length-contraction result, is explicitly only a sketch: two concentration steps in Appendix A.3 are stated without proof. That is not a fatal flaw, because the proposition is not used in Theorem 1 or Theorem 2 and the paper labels it a sketch, but it is presented as a theoretical result and should be either completed or formally downgraded to a conjecture with empirical support before the paper is treated as having fully discharged its theoretical claims. There is also a minor overstatement in Section 5.1: Table 2 says every entry is within Monte Carlo error of 0.90, but the t3 entry of 0.963 is more than three standard errors above nominal; this is a wording slip, not a substantive issue. The frozen-score requirement in Section 3.3 is a genuine limitation, but the paper states it clearly, and the price of type-adaptivity is honestly quantified.\n\nThis paper deserves a serious referee. I would send it out and ask for the Proposition 1 proof to be completed or its status downgraded, plus the Table 2 sentence fixed. The main contribution is solid, the simulations support it, and the invariance theorem is a genuine step forward for practitioners who need uncertainty sets on sensor streams under unknown marginal transformations.","headline":"Rank-invariant conformal changepoint localization is a genuinely useful idea; the coverage and invariance theorems are sound, and the only real blemish is an explicitly unfinished proof of the length-contraction proposition.","tokens_in":15743,"tokens_out":1117,"would_cite":true,"duration_ms":13775,"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":"Rank-only scores make changepoint confidence sets transform-proof","keywords":["changepoint localization","conformal inference","rank statistics","confidence set","distribution-free","invariance","efficiency transfer","finite-sample coverage"],"falsifier":"Run ARC on a fixed window and again on the same window transformed by $g(x)=e^x$, sharing the permutation seed: the paper predicts identical p-values at every candidate, so any observed difference would disprove Theorem 2.","tokens_in":1646,"feed_emoji":"🎯","tokens_out":2269,"duration_ms":71203,"temperature":0.7,"pith_summary":"The paper proposes ARC, a family of changepoint-localization scores that read the data only through within-segment ranks. The aim is to keep the finite-sample coverage guarantee of conformal localization while making efficiency, meaning the length of the reported confidence set, transferable across distributions. ARC claims two properties: every frozen weight configuration, including a randomly initialized or mistrained network, yields coverage at least $1-\\alpha$; and the entire confidence set is almost surely unchanged under strictly increasing transforms of the data, so set-length distributions certified once hold across the whole monotone orbit of a distribution pair. If true, users can certify set lengths on one distribution and carry them verbatim to heavy-tailed or skewed re-expressions, whereas plug-in likelihood-ratio scores see their set lengths drift and inflate.","feed_headline":"Rank-only scores make changepoint confidence sets transform-proof","feed_subtitle":"Coverage stays finite-sample for any frozen score; set length holds across every monotone re-expression.","key_machinery":"The machinery is a bank of rank-only segment statistics: a rank-CUSUM location channel built from maximized Wilcoxon evidence, a Mood-type squared-rank scale channel, fixed maximum or convex combinations of the two channels, and a single-hidden-layer network fed by downsampled rank paths, trained on synthetic data and frozen. Each candidate split $t$ is scored by the maximum of the relevant channel applied to the two segments, and a p-value is computed by within-segment permutation; the confidence set collects all $t$ whose p-value exceeds $\\alpha$. Rank invariance carries the argument, since only the rank vector enters the score, and the frozen-parameter requirement keeps the permutation p-value exactly super-uniform.","core_discovery":"ARC's central discovery is an efficiency-transfer theorem: because every ARC statistic is a function of within-segment ranks, applying any strictly increasing $g$ to the data leaves every rank, every split statistic, and every permutation p-value unchanged, so $C_{1-\\alpha}(g(X))=C_{1-\\alpha}(X)$ almost surely under shared permutation randomness. Consequently the joint law of the confidence set and changepoint depends on the pre- and post-change distributions only through their rank structure, and lengths certified once hold verbatim for every monotone-equivalent pair, however heavy-tailed. This contrasts with plug-in likelihood-ratio scores, whose set length changes with every re-expression. The paper also states finite-sample coverage for any frozen score and shows that the price of invariance is bounded by classical rank-test efficiency results.","pith_inferences":["If the transfer theorem is correct, localization-set calibration could become a table lookup: certify lengths once per rank structure and reuse them across every monotone-equivalent data pair, potentially removing per-dataset tuning of set sizes.","The same rank-invariance argument may extend to online changepoint localization built on e-values or exchangeability martingales, giving anytime-valid inference that is also invariant under monotone rescalings; the paper does not claim this extension.","A natural testable extension is per-coordinate rank channels for multivariate windows, turning type-awareness into coordinate attribution without forfeiting invariance; the paper gestures at but does not develop this direction.","The well-log results suggest ARC's practical benefit is insurance rather than uniform improvement: it is slightly longer than a tuned plug-in at the distribution the plug-in was built for, but it protects against distribution shift."],"forward_implications":["A set-length distribution certified once, for example on Gaussian simulations, transfers verbatim to every monotone-equivalent pair: lognormal, logistic-compressed, or arbitrarily heavy-tailed images of the same rank experiment.","Plug-in likelihood-ratio scores retain coverage but their set lengths inflate under re-expression, whereas ARC set lengths are unchanged to the last replication across the tested monotone transforms.","No deployment-side calibration is needed for validity: every frozen weight state, including random initialization and a network trained on permuted labels, keeps coverage at the nominal level, with only set length affected.","On the well-log benchmark, combined ARC scores localize consensus annotations within three to five candidates, and an empty set is interpreted as a model-misfit signal at level $\\alpha$ rather than a failure.","Under AR(1) serial dependence, i.i.d. within-segment permutations can drop coverage to about 0.30, and circular block permutations only repair it approximately to 0.84–0.92, so exact coverage is confined to within-segment exchangeable data."],"supporting_citations":[{"why":"Supplies the conformal changepoint-localization backend and the universality result that every valid procedure is an instance of the wrapper with some score.","marker":"[14]"},{"why":"Provides the matrix-of-p-values conformal localization construction and the conformal Neyman-Pearson lemma identifying the oracle likelihood-ratio score.","marker":"[6]"},{"why":"Establishes exactness of Monte Carlo permutation p-values, the property that makes coverage hold for any frozen ARC score.","marker":"[13]"},{"why":"Positions rank procedures as the maximal invariants under the monotone transformation group, which underlies the efficiency-transfer claim.","marker":"[12]"},{"why":"Bounds the asymptotic relative efficiency of rank statistics against parametric competitors, used to state the price of invariance.","marker":"[4]"},{"why":"Defines the Wilcoxon rank-sum statistic that forms the ARC location channel.","marker":"[39]"},{"why":"Defines the Mood squared-rank statistic that forms the ARC scale channel.","marker":"[26]"},{"why":"Provides the CUSUM scanning device that turns segment statistics into split statistics over candidate changepoints.","marker":"[30]"},{"why":"Supplies the well-log benchmark data and human annotations used in the real-data localization protocol.","marker":"[35]"}],"fun_headline_variants":["Rank-only scores: transform-proof changepoint sets","Set length invariant under any monotone re-expression","Efficiency transfer: one certification, all monotone pairs","One rank-based set, all monotone equivalents"],"cache_read_input_tokens":17920,"weakest_assumption_plain":"The guarantee rests on the score and any learned parameters being fixed before the test window is seen, and on the data inside each segment being exchangeable; if the combination rule is chosen after viewing the window, or if serial dependence breaks within-segment exchangeability, the coverage argument no longer holds exactly.","fun_headline_variants_meta":{"raw":{"variants":["Rank-only scores: transform-proof changepoint sets","Set length invariant under any monotone re-expression","Efficiency transfer: one certification, all monotone pairs","One rank-based set, all monotone equivalents"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001784,"raw_usage":{"total_tokens":7065,"prompt_tokens":1011,"completion_tokens":6054,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":627,"completion_tokens_details":{"reasoning_tokens":5992}},"tokens_in":627,"tokens_out":6054,"duration_ms":45398,"temperature":1.0,"reasoning_tokens":5992,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:35:55.474869+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run ARC on a fixed window and again on the same window transformed by $g(x)=e^x$, sharing the permutation seed: the paper predicts identical p-values at every candidate, so any observed difference would disprove Theorem 2.","supporting_citations":[{"cited_title":"Offline changepoint localization using a matrix of conformal p-values.Transactions on Machine Learning Research, 2026","cited_arxiv_id":null,"evidence_quote":"Provides the matrix-of-p-values conformal localization construction and the conformal Neyman-Pearson lemma identifying the oracle likelihood-ratio score."},{"cited_title":"Academic Press, New York, 1967","cited_arxiv_id":null,"evidence_quote":"Positions rank procedures as the maximal invariants under the monotone transformation group, which underlies the efficiency-transfer claim."}],"review_version":1}