{"id":"23cf266e-026c-457c-804b-dcf839d2dd60","arxiv_id":"2603.08994","paper_version":5,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Known Mersenne prime exponents are enriched for large τ(p−1), and a cyclotomic-layer heuristic encodes this as P(Mp prime) ∝ (log p)^{S(p)}/p.","lead":"This paper reports that known Mersenne prime exponents tend to have more divisors in p−1 than nearby primes of the same size, and proposes a heuristic model in which this divisor structure slightly boosts the chance that 2^p−1 is prime. The proposed correction is finite-scale and does not change the classic Wagstaff asymptotic estimate.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The S(p) enrichment may be a proxy for the known p≡1 mod 4 bias in Mersenne exponents; controls are not stratified by residue, so the central empirical claim is not yet isolated.","rationale":"The reader's weakest_assumption focused on the model's proportionality L≈kτ(p−1) and its extrapolation beyond p<10^6. That is a real concern, but the more load-bearing vulnerability is upstream: the empirical enrichment itself may be confounded by p mod 4. Since S(p) is built from τ(p−1), and τ(p−1) depends on v2(p−1), the known Wagstaff residue effect (p≡1 mod 4 favored) can produce exactly the kind of S-elevation the paper reports. The control procedure does not stratify by residue, so the central empirical claim is not yet cleanly isolated. This is not an objection to the heuristic nature of the model; it is a specific, testable control omission. The proposed test—stratifying all principal analyses by p mod 4—would settle whether the effect is new or already known. Until then, the reader's CONDITIONAL verdict remains appropriate; no change to the verdict is needed, but the condition should include this stratification check.","tokens_in":16483,"tokens_out":15649,"duration_ms":140851,"concrete_test":"Recompute the principal statistics (mean S, median π_W, Wilcoxon signed-rank, and the stratified permutation test) separately for p≡1 mod 4 and p≡3 mod 4, with control sets restricted to the same residue class. Equivalently, add p mod 4 as an extra stratum in the conditional logistic model of §A.3 and in the permutation test of §A.4. If the S-elevation and percentile-rank inflation vanish or fall below significance within both residue strata, the observed enrichment is explained by the known Wagstaff mod-4 bias and the new divisor-structure claim fails. If significant enrichment survives in both strata, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's empirical claim—that Mersenne prime exponents have elevated S(p) relative to nearby primes—is confounded by p mod 4. For p≡3 mod 4, v2(p−1)=1, so τ(p−1)=2τ((p−1)/2); for p≡1 mod 4, v2(p−1)≥2, giving an extra factor of at least 3/2 in τ(p−1). Thus S(p)=logτ(p−1)/loglogp is systematically larger for p≡1 mod 4. The paper itself, in §4.1, recalls Wagstaff's refinement that Mersenne primes are more probable for p≡1 mod 4, and Table 1 indeed shows a 31/17 split toward p≡1 mod 4. Appendix A's control windows are not stratified by residue class, so a Mersenne exponent with p≡1 mod 4 is often compared against nearby p≡3 mod 4 controls with structurally lower S. The observed mean shift (1.316 vs 1.138) and inflated percentile ranks could therefore be largely a residue-class artifact rather than evidence for the new divisor-structure mechanism. The proposed law P∝(logp)^S/p also omits the mod-4 factor and may simply be absorbing the known Wagstaff residue effect into S. If the enrichment disappears after conditioning on p mod 4, the central empirical claim has no independent support.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper reports an empirical bias in the divisor structure of p−1 for known Mersenne prime exponents. It defines S(p)=log τ(p−1)/log log p and compares S(p) for the 48 known exponents p≥13 against nearby prime controls. The authors report moderate enrichment (mean percentile ~0.65, Cohen's d≈0.56) with p-values ≈0.001–0.002 across sign, Wilcoxon, KS, conditional logistic, and permutation tests. To interpret this, they develop a heuristic model based on cyclotomic divisors of 2^{p−1}−1, aggregating congruence constraints to propose P(M_p prime|S)≈C (log p)^{S(p)}/p, which they claim reproduces the observed S>1 versus S<1 imbalance while preserving the Wagstaff scale. The paper is explicitly framed as heuristic and finite-scale.","tokens_in":16944,"tokens_out":14144,"duration_ms":112631,"significance":"Strengths: the statistical analyses are carefully described and reproducible, using distribution-free procedures and stratified permutation tests; the paper is honest about limitations. If, after proper conditioning, the enrichment survives, it would be an interesting secondary arithmetic signal in the distribution of Mersenne exponents. Weaknesses: the central empirical claim is currently confounded by the known p≡1 mod 4 bias, which mechanically inflates S(p); the model is fitted on the same data and extrapolated beyond its measured range. The paper's contribution therefore hinges on whether the enrichment persists within residue classes.","major_comments":[{"comment":"The control sets are not stratified by p mod 4. For p≡3 mod 4, p−1=2·odd, so τ(p−1)=2τ(odd); for p≡1 mod 4, p−1 has v2≥2, so τ(p−1)≥3τ(odd′). Hence S(p)=logτ/loglogp is systematically larger for p≡1 mod 4. Table 1 shows a 31/17 bias of the known Mersenne exponents toward p≡1 mod 4 (the Wagstaff residue effect recalled in §4.1). Since the controls in A.1 include both residues, the reported sign test (p=0.0021), Wilcoxon (p=0.0014), KS (p=0.00046), and permutation (p=0.0012) may reflect this known residue effect rather than a new divisor-structure signal. The analyses must be rerun with controls restricted to the same residue class (and/or with p mod 4 as a covariate in the conditional logistic model).","section":"Appendix A.1–A.4 and §3"},{"comment":"The key scaling L(p)≈kτ(p−1) is measured only for p<10^6. The entries for p∼10^7 and 10^8 (k≈0.684, θ≈0.82) are explicitly described as 'finite-scale projections' rather than direct factor-enumeration results. Nevertheless, Table A.7 uses these projected values to generate predicted counts for p<10^7 and p<10^8, including the claimed agreement with the observed 28/6 and 37/10 splits. Because this extrapolation is load-bearing for the largest-range claims, the authors should either restrict the model to the directly computed range or supply independent evidence for the proportionality beyond p=10^6.","section":"Appendix A.6–A.7"},{"comment":"The proposed law P∝(logp)^S/p is introduced after the empirical enrichment is observed, and its parameters (k, w_2, a_trunc, ε) are calibrated on the same Mersenne-prime data. The matched 'prediction' in A.7 is therefore an in-sample consistency check, not a falsifiable out-of-sample prediction. To claim predictive power, all parameters should be fixed using exponents below a threshold (or a random subset) and then evaluated on held-out exponents. As it stands, the agreement between the model and the observed S>1/S<1 split is expected by construction.","section":"§5.4 and A.7"},{"comment":"The refined law P(M_p prime|S)≈C(logp)^S/p omits the mod-4 factor that Wagstaff's refinement (§4.1) requires. Since S(p) is mechanically higher for p≡1 mod 4, the model may simply be re-encoding the established residue bias as a 'structural' effect. The paper should show that the S-effect is not a proxy for mod 4, e.g., by demonstrating the enrichment within each residue class, or by including both S(p) and p mod 4 in the conditional model and reporting the partial effect of S.","section":"§4.2–§5.5"}],"minor_comments":[{"comment":"No equation numbers are used, which makes precise reference cumbersome; adding numbers would improve readability.","section":"General"},{"comment":"The statement that log τ(p−1) 'fluctuates on a scale comparable to log log p' is informal; a reference to known distribution results for τ(n) would help.","section":"§2.4"},{"comment":"The upper limit t in the sum ∑_{n=2}^t is never defined; it should be specified as the maximum factorization length.","section":"§5.3"},{"comment":"The caption calls π_{1000} and π_{5000} 'percentile ranks' but the values are proportions in [0,1]; clarify the intended scale.","section":"Table 1"},{"comment":"The sentence 'Exponents below p=13 are excluded' could explicitly state that the four excluded Mersenne exponents are 2, 3, 5, and 7.","section":"A.1"},{"comment":"The regression table presents 'Pearson r, Spearman ρ, k 95%CI' in a single column; separate columns would be clearer.","section":"A.6"}],"recommendation":"major_revision","confidential_remarks":"The residue-class confound is the main risk. If a reanalysis with residue-matched controls removes the effect, the central claim collapses. The editor may wish to ask specifically for that reanalysis before further consideration. Also, the in-sample 'prediction' and the extrapolation beyond p<10^6 should be addressed; these are fixable but require new computations."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nPunchline: the empirical enrichment of S(p) among Mersenne prime exponents is probably real, but as presented it might just be the known p≡1 mod 4 bias wearing a new hat. The paper is honest and careful, but its central claim is not isolated from residue class, and the final probability law is a curve fit, not a prediction.\n\nWhat's actually new: the S(p) statistic (log τ(p−1)/log log p) is a clean normalization, and the cyclotomic-layer interpretation is a genuine attempt to give the divisor structure a mechanism. The statistical work is solid as far as it goes: sign test, Wilcoxon, KS, conditional logistic, permutation tests, all with p-values around 0.001. The authors are transparent that this is heuristic and finite-scale.\n\nThe soft spot is the p mod 4 confound. For p≡3 mod 4, v2(p−1)=1, so τ(p−1)=2τ(odd). For p≡1 mod 4, v2(p−1)≥2, giving an extra factor at least 3/2. So S(p) is systematically larger for p≡1 mod 4. Wagstaff's refinement says Mersenne primes are likelier when p≡1 mod 4, and Table 1 shows a 31/17 split. The control windows in Appendix A mix residue classes, so a 1 mod 4 exponent is compared to a mix that tilts toward 3 mod 4 controls with structurally lower S. The paper never reports a residue-stratified comparison. If the enrichment vanishes after conditioning on p mod 4, the central empirical claim has no independent support.\n\nThe explanatory model in §§2–5 is speculative but not crazy. The cyclotomic decomposition is real, but the step from algebraic identities to 'modular filters' is a heuristic leap. The law P∝(log p)^{S(p)}/p is constructed after seeing that exponents have high S, and the exponent S(p) enters through a fitted coefficient k in L≈k τ(p−1). Appendix A.7's 'prediction' is a re-description of the input, not an out-of-sample test. The paper itself admits this, which is to its credit.\n\nVerdict: this deserves a serious referee. The empirical claim is specific and testable, and the statistical toolkit is respectable. A referee should ask for a residue-stratified control analysis and a genuine out-of-sample or pre-registered test. Without that, don't take the enrichment at face value. I wouldn't cite it as evidence yet, but I'd read the revision.","headline":"The S(p) enrichment is plausibly real but confounded by the known p≡1 mod 4 bias; the final law is a curve fit, not a prediction—worth refereeing, but only as a conditional accept.","tokens_in":17337,"tokens_out":3055,"would_cite":false,"duration_ms":24711,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["11A41","11N05","11Y11","62G10"],"pacs":[],"model":"deepseek-v4-flash","headline":"Known Mersenne prime exponents carry a secondary arithmetic signal: divisor-rich p-1 appears more often than exponent size alone predicts.","keywords":["Mersenne primes","divisor function","cyclotomic polynomials","arithmetic bias","prime distribution","probabilistic number theory","experimental number theory"],"falsifier":"Record the S(p) values of the next Mersenne prime exponents discovered beyond 136,279,841; if they are not predominantly above the local median of S among nearby primes, the bias claim fails. Alternatively, partially factor A_p = (2^{p-1} − 1)/p for p in 10^6–10^7 and check whether ω_{<p}(A_p)/τ(p-1) stays near 0.684; a sharp drop would invalidate the layer-counting step.","tokens_in":16385,"feed_emoji":"🔢","tokens_out":5088,"duration_ms":41595,"temperature":0.7,"pith_summary":"This paper claims that the distribution of known Mersenne prime exponents is not fully explained by exponent size. It introduces S(p) = log τ(p-1) / log log p and shows that known exponents, excluding the smallest cases, sit systematically higher in S(p) than nearby prime controls of comparable size. To explain this, the paper builds a heuristic model in which divisors of p-1 create cyclotomic layers that filter candidate factorizations of 2^p − 1, leading to a refined probability law P(M_p prime | S) ≈ C (log p)^{S(p)} / p. If correct, the classical size-only heuristic remains the asymptotic law, but a finite-scale arithmetic bias tied to the divisor structure of p-1 is real and measurable.","feed_headline":"Mersenne prime exponents skew toward divisor-rich p-1","feed_subtitle":"Known exponents show a measurable enrichment in S(p) after matching size, hinting at a secondary arithmetic bias.","key_machinery":"The normalized divisor parameter S(p) = log τ(p-1) / log log p; the cyclotomic identity 2^{p-1} − 1 = ∏_{d|p-1} Φ_d(2); the two-factor diagonal residue classes that concentrate all prime factors of M_p into one residue class modulo a; and the empirical scaling ω_{<p}(A_p) ≈ k τ(p-1) with k ≈ 0.684, which converts divisor count into effective filter count.","core_discovery":"On the paper's own terms: Mersenne prime exponents have elevated divisor complexity of p-1 relative to local prime controls, and this elevation survives matching by exponent size. The mean percentile rank of S(p) in a 5000-prime window is 0.648 versus a null expectation of 0.5; sign, Wilcoxon, and permutation tests reject the null with p-values around 0.002. The paper interprets this as evidence for a structural refinement of the standard probability estimate for M_p being prime, preserving the log p / p scale after averaging over S.","pith_inferences":["The finite sample of 48 known exponents makes the statistical signal fragile; the strongest confirmation would come from the next discovered exponent, not from reanalysis of the same data.","If the bias persists, S(p) becomes a cheap prefilter for computational searches; if it does not, the effect is likely a selection artifact of how the known set was assembled.","The k ≈ log 2 saturation argument suggests the layer-density coefficient may be bounded by a capacity effect, which would make the bias weaker or saturating at very large p."],"forward_implications":["If the bias is real, size-only estimates under-predict the probability for exponents whose p-1 has many divisors and over-predict for sparse ones, without changing the average behavior.","Search strategies for new Mersenne primes could rank candidate exponents by S(p) within a given size range.","The model yields testable predictions: the split of new exponents between S(p) > 1 and S(p) < 1 should remain skewed rather than approach 50/50.","The empirical k ≈ 0.684 relation gives a direct way to estimate the number of effective modular filters from τ(p-1) alone."],"fun_headline_variants":["Mersenne primes skew to divisor-rich p-1","Mersenne exponents show divisor-structure bias","p-1 divisor complexity hints at Mersenne bias","Mersenne prime exponents favor high-divisor p-1","Divisor structure of p-1 predicts Mersenne primality"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The model stands on the premise that the number of cyclotomic layers that actually constrain factorizations is proportional to τ(p-1) (L ≈ k τ(p-1)), a proportionality measured only below p ≈ 10^6 and assumed to continue at larger scales.","fun_headline_variants_meta":{"raw":{"variants":["Mersenne primes skew to divisor-rich p-1","Mersenne exponents show divisor-structure bias","p-1 divisor complexity hints at Mersenne bias","Mersenne prime exponents favor high-divisor p-1","Divisor structure of p-1 predicts Mersenne primality"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000207,"raw_usage":{"total_tokens":1282,"prompt_tokens":838,"completion_tokens":444,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":582,"completion_tokens_details":{"reasoning_tokens":362}},"tokens_in":582,"tokens_out":444,"duration_ms":4303,"temperature":1.0,"reasoning_tokens":362,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T18:30:53.422262+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record the S(p) values of the next Mersenne prime exponents discovered beyond 136,279,841; if they are not predominantly above the local median of S among nearby primes, the bias claim fails. Alternatively, partially factor A_p = (2^{p-1} − 1)/p for p in 10^6–10^7 and check whether ω_{<p}(A_p)/τ(p-1) stays near 0.684; a sharp drop would invalidate the layer-counting step.","supporting_citations":[],"review_version":2}