{"id":"08119a4f-7700-4e8b-b791-a72d6d30fa9a","arxiv_id":"2606.20859","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":3,"one_line_summary":"Shifted Legendre polynomials turn Simple Jumper martingales into multi-moment exchangeability tests, with a mean-field variational form that avoids exponential state-space cost.","lead":"The paper builds conformal test martingales that bet on higher-order moments of p-values using shifted Legendre polynomials, so they can catch variance and skewness shifts that the classic Simple Jumper misses. A mean-field variational version keeps the cost linear instead of exponential, making multi-moment online exchangeability testing practical.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified beyond the paper's own open-problem caveat on mean-field power loss.","rationale":"The paper's strongest claim is that VLJ is a valid, linearly scaling CTM that empirically matches PLJ capital growth under multi-moment shifts, with the Composite form recommended as default. Validity follows immediately from the construction: consensus parameters are F_{n-1}-measurable, the product betting function integrates to 1 after the explicit Z normalisation, and non-negativity is guaranteed by the standard grid lying inside [-1,1] (Section 7). The open theoretical gap on approximation quality is already disclosed, so it does not constitute an unacknowledged load-bearing flaw. Empirical support is narrow but internally consistent and reproducible via the released package. The reader's CONDITIONAL verdict with high confidence therefore already reflects the correct balance; no adjustment is warranted.","tokens_in":15684,"tokens_out":458,"duration_ms":4395,"concrete_test":"Independently re-implement Algorithm 4 for K={1,2,3} on the same Wine Quality orderings (or any public conformal p-value stream) and verify that the final log10 Mn of VLJ stays within 5% of PLJ while using only 15 states; if the gap exceeds that under a pure higher-moment alternative (e.g., Beta(0.3,0.3)), the 'minimal power loss' claim would need stronger qualification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader's weakest_assumption correctly flags that formal error bounds on the mean-field approximation remain open (Section 9) and that empirical support for VLJ≈PLJ is limited to one UCI stream plus synthetic betas. That is a genuine limitation on treating VLJ as a drop-in default without caveat, but it is already stated by the authors and does not undermine the central validity claim: the process is a valid CTM by predictability of the consensus parameters (Algorithm 4 and the paragraph preceding it), the grid is proven safe for all degrees (Section 7), and the linear-cost construction is correctly derived. No hidden inconsistency or unacknowledged failure mode appears in the argument itself.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper generalises the Simple Jumper conformal test martingale by replacing its linear betting function with shifted Legendre polynomials. It defines the Simple Legendre Jumper (arbitrary single degree k), the Product Legendre Jumper (product over a set K of degrees, with explicit normalisation Z(ε) for higher-order cross terms), and the Variational Legendre Jumper (mean-field factorisation into independent sub-jumpers whose wealth-weighted consensus parameters are plugged into the product bet). A Composite variant averages over a grid of jumping rates. Validity under exchangeability follows from the integral constraint and predictability; non-negativity of the standard grid E for all degrees is established via |P_k|≤1. Empirical trajectories on synthetic beta alternatives and the Wine Quality dataset (four orderings) show that multi-degree PLJ/VLJ accumulate more capital than single-degree SLJ under multi-moment shifts, that VLJ closely tracks PLJ at linear rather than exponential cost, and that the composite form supplies a wealth floor and automatic rate adaptation.","tokens_in":15890,"tokens_out":1008,"duration_ms":13346,"significance":"If the constructions hold, the work supplies a practical, distribution-free tool for online multi-moment exchangeability testing that removes the Simple Jumper’s restriction to location shifts while remaining computationally feasible. The explicit link to Neyman’s smooth test, the closed-form treatment of Z(ε), the proof that the standard grid is safe for every degree, and the open-source implementation in online-cp are concrete strengths. The variational reduction of the “jumping tax” is a useful engineering contribution even without formal power bounds, and the composite recommendation is immediately usable by practitioners who do not know the shift timescale.","major_comments":[{"comment":"Section 9 and the empirical support for VLJ: the paper correctly flags that formal error bounds on the mean-field approximation remain open. The claim of “minimal power loss” (abstract, §5.1) rests on a single UCI stream (Wine Quality, four orderings) plus a handful of synthetic beta mixtures (Figs. 5–7). For |K|≥3 the approximation quality is therefore uncharacterised beyond these examples. Either additional controlled alternatives that isolate higher-order moments, or a quantitative statement of the observed relative log-capital gap, would be needed before VLJ can be presented as a drop-in default without caveat.","section":null},{"comment":"Section 8 / Table 3: only one real-world dataset and a fixed nonconformity measure (1-NN ratio) are used. The hierarchy PLJ/VLJ ≫ SLJ(k=1) ≫ SLJ(k=2,3) may be specific to the red/white wine change-point structure. At least one additional stream (or a synthetic multi-moment change-point with known ground-truth moments) would strengthen the claim that multi-degree betting is systematically superior under realistic distributional shift.","section":null}],"minor_comments":[{"comment":"Abstract vs. body: the abstract states “constant time per step”; Algorithm 4 and §5.1 correctly give O(|K|·g). Align the wording.","section":null},{"comment":"Figure 2 caption still reads f^{(1)}_ε while the plot is for k=2; likewise a few “decreses” / “comaratively” typos remain.","section":null},{"comment":"Section 5: the derivation of Z(ε) for |K|≥3 is clear for K={1,2,3,4}, but a short remark on how the lookup table is built for larger K (or a pointer to the package) would help reproducibility.","section":null},{"comment":"Notation: the same symbol E is used both for the ε-grid and for expectation; a brief disambiguation would avoid momentary confusion.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The central validity argument is clean and the open-problem caveat on mean-field power is already stated by the authors; the two major comments are therefore requests for stronger empirical grounding rather than corrections of inconsistency. The manuscript is a natural fit for a methods-oriented ML/statistics venue; I would not block acceptance once the empirical caveats are made more explicit and the abstract wording is aligned."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is a solid methods note that does exactly what it claims. The new pieces are the Product Legendre Jumper (with the explicit Z(ε) normalisation for |K|≥3), the named jumping tax, and especially the Variational Legendre Jumper that factorises the product into independent sub-jumpers plus a consensus product bet. That last construction is the real algorithmic contribution: linear cost in the number of degrees while remaining a valid conformal test martingale by predictability of the consensus parameters.\n\nWhat the paper does well is keep the foundations tight. Martingale validity follows immediately from the integral constraint and the fact that the betting functions are determined before the next p-value. Non-negativity of the standard grid for every degree is proved cleanly via |P_k|≤1. The cross-term coefficients for small K are derived correctly. The composite over jumping rates is the sensible default and inherits the wealth floor from the J=1 component. Code is already in online-cp, which is more than most notes deliver.\n\nThe soft spots are real but already flagged by the author. Formal power-loss bounds for the mean-field step are left open (Section 9), and the empirical case that VLJ tracks PLJ rests on one UCI stream (Wine Quality under four orderings) plus a few synthetic betas. That is enough to show the idea works and that single-degree SLJs of order 2 or 3 alone can miss the dominant location shift, but it is not a broad stress test. Free parameters (grid, J, K) are the usual ones; nothing is hidden. Citation pattern is appropriate: Vovk for the jumper infrastructure, Neyman for the historical Legendre motivation.\n\nThis is for people who already run conformal test martingales and want multi-moment power without exponential state spaces. It will not reorganise the field, but it is immediately usable and formally grounded. I would send it to peer review; a referee can push for more datasets and a start on the approximation bounds, but the core is ready.","headline":"Clean, usable multi-moment extension of the Simple Jumper with a practical linear-cost variational form; math holds, empirics are narrow but honest.","tokens_in":16515,"tokens_out":520,"would_cite":true,"duration_ms":5957,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62L10","62G10","60G42"],"pacs":[],"model":"grok-4.5","headline":"Shifted Legendre polynomials turn the Simple Jumper into a scalable martingale that bets on variance, skewness and higher moments of conformal p-values.","keywords":["conformal test martingales","exchangeability testing","shifted Legendre polynomials","Simple Jumper","distribution shift","mean-field approximation","online monitoring"],"falsifier":"Run both the Product and Variational Legendre Jumpers with the same multi-degree set K on a controlled multi-moment shift (e.g., a known Beta or mixture sequence) and check whether the final log-wealth of the variational version falls more than a few percent below the exact product version; a large, systematic gap would falsify the “minimal power loss” claim.","tokens_in":16545,"feed_emoji":"📈","tokens_out":961,"duration_ms":8521,"temperature":0.7,"pith_summary":"Exchangeability—the idea that the future looks like the past—is routinely violated by distribution shift, yet many online tests only notice when the mean of conformal p-values moves. This paper shows that the familiar Simple Jumper can be rewritten with shifted Legendre polynomials of any degree, so a single betting function can target variance, skewness or kurtosis in isolation. Combining several degrees into a product betting function captures multi-moment shifts at once, but the joint state space grows exponentially and dilutes capital—the “jumping tax.” A mean-field (variational) factorisation runs one cheap sub-jumper per degree, forms a consensus parameter vector, and places a single product bet; the resulting process remains a valid conformal test martingale, costs linear time, and empirically matches the exact product’s capital growth. Averaging over a small grid of jumping rates further supplies a wealth floor and automatic adaptation to unknown shift timescales. On a real wine-quality classification stream the multi-degree versions accumulate substantially more evidence against exchangeability than any single-degree martingale, while staying quiet under pure permutation.","feed_headline":"Martingales that bet on variance and skewness of p-values","feed_subtitle":"Legendre Jumpers detect multi-moment shifts online while keeping false-alarm control and linear cost","key_machinery":"The Variational Legendre Jumper: independent Simple Legendre sub-jumpers (one per degree) produce wealth-weighted consensus parameters; a single normalised product of the corresponding Legendre betting functions is then used as the global bet, guaranteeing a valid test martingale whose cost scales linearly in the number of degrees.","core_discovery":"A family of conformal test martingales built from shifted Legendre polynomials—Simple, Product, Variational and Composite—extends the Simple Jumper from mean shifts to simultaneous higher-order moment deviations while preserving the martingale property and, in the variational case, reducing per-step cost from exponential to linear with negligible empirical power loss.","pith_inferences":["The same mean-field factorisation could be applied to other orthogonal polynomial families or to sleeping/waking jumper architectures already used in conformal testing.","A sequential model-selection rule that grows or shrinks the active degree set K on the fly would mirror classical data-driven smooth tests while staying inside the martingale framework.","If formal approximation bounds can be obtained via variational divergence techniques, the method would supply explicit power guarantees for |K|≥3 rather than relying solely on empirical match."],"forward_implications":["Practitioners can monitor real-time conformal streams for variance or skewness collapse without waiting for a mean shift.","The Composite form (average over a grid of jumping rates) becomes the default online exchangeability test when the shift timescale is unknown.","Higher-degree Legendre bets remain valid on any conformal p-value sequence, so the same code can be dropped into existing online conformal pipelines.","The linear-cost variational construction makes it feasible to keep many polynomial degrees active simultaneously rather than committing to a single order in advance."],"fun_headline_variants":["Betting on higher moments with Legendre Jumper martingales","Legendre Jumpers detect variance and skewness shifts online","Polynomial betting extends Simple Jumper to multi-moments","Variational Legendre Jumper scales higher-order testing","Conformal martingales that bet on p-value moments"],"cache_read_input_tokens":128,"weakest_assumption_plain":"That the mean-field consensus parameters keep essentially the same detection power as a fully joint product chain for the degrees people actually use, even though the paper leaves formal error bounds open and shows the claim on one real data set plus synthetic betas.","fun_headline_variants_meta":{"raw":{"variants":["Betting on higher moments with Legendre Jumper martingales","Legendre Jumpers detect variance and skewness shifts online","Polynomial betting extends Simple Jumper to multi-moments","Variational Legendre Jumper scales higher-order testing","Conformal martingales that bet on p-value moments"]},"model":"grok-4.5","effort":"low","cost_usd":0.00564,"raw_usage":{"total_tokens":1502,"prompt_tokens":795,"num_sources_used":0,"completion_tokens":65,"cost_in_usd_ticks":56400000,"prompt_tokens_details":{"text_tokens":795,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":642,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":795,"tokens_out":65,"duration_ms":5105,"temperature":1.0,"reasoning_tokens":642,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T17:40:47.854392+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Run both the Product and Variational Legendre Jumpers with the same multi-degree set K on a controlled multi-moment shift (e.g., a known Beta or mixture sequence) and check whether the final log-wealth of the variational version falls more than a few percent below the exact product version; a large, systematic gap would falsify the “minimal power loss” claim.","supporting_citations":[],"review_version":2}