# The delta mechanism and x = 37 through the HL lens: the second singular-series layer IS the residual sub-field, delta itself survives all three candidate families, and 37 stays a G2-side anomaly no comb explains *(2026-08-21. Staging note; nothing here is integrated into a live document. Producer, formally embedded: `research/attack-delta37-01.js` (0.6 s, no sieving — every measured number parsed from committed embeds; code-sha256 b134f6ad…, out-sha256 b2b5909b…). Calibration marked on every claim: PROVEN, VERIFIED by exact computation, MEASURED, PREDICTED, REFUTED, NOT FOUND. Brief: with W1 identified as the HL twin-twin comb (`w1-singular-series.md`), score three candidate mechanisms for the depth term exp(−δ·θ/m̄) of `mp-derivation.md` — each on its parameter-free prediction FIRST, with the far-tail flattening (0.277 → 0.252) as the discriminating data — and read x = 37 through the same lens.)* ## 0. The answer **(a) δ is not sourced by any of the three candidate families.** The scored verdicts, each parameter-free before any refit (a δ source needs tilt slope −0.277 per unit z; the measured flattening is φ = +0.0247): | candidate | predicted tilt (train) | predicted φ | verdict | |---|---|---|---| | (iii) finite-size / edge, derived exactly | −0.0024 | +0.0021 | **REFUTED**, 2 orders short (window-edge ≤ 2.5·10⁻⁸ pooled; W1-cutoff tail ≤ 2.0·10⁻²) | | (i-a) published second moments of 𝔖 | 0 | 0 | **no channel**: θ_p is deterministic and W1 is carried exactly; an ensemble variance prices spread, not a per-fold mean | | (ii) the derived thinning null's own depth structure | −0.0600 | +0.0764* | right sign, **22% of δ**; custody-licensed (reproduces THN's zero-parameter window totals: 5481.9) | | (i-b) the interior 6-tuple comb (second 𝔖 layer), exact | −0.0208 | +0.0241 | **8% of δ — but it is the residual sub-field, §2** | *(φ entries are band-OLS summaries of the prediction; under the record's own estimators — train OLS + deep Poisson MLE, the only fair comparison — see §3.)* **But two of the losers, combined, derive the residual sub-field that `mp-derivation.md` §5 named and left "deterministic-looking and NOT derived" — with zero new parameters.** Crediting the exact thinning-null depth structure (ii) plus the exact interior comb (i-b) into the D4 model and refitting only the law's own two scalars: - train roughness **sd 0.1715 → 0.0834** (the interior comb alone: 0.0830; Pearson correlation of the interior comb with the D4 train residual field **0.907**, Spearman 0.842, over the 37 train folds) [MEASURED]; - the far-tail flattening collapses **φ = 0.0247 → 0.0085** (δ′ = 0.1964 train / 0.1879 deep — nearly flat) [MEASURED]; - **all four flagged folds drop below 3**: 311: 3.10 → 2.63, 331: 3.21 → 2.22, 409: 3.28 → 2.31, 631: 3.36 → 2.45; deep |pull| > 3 count 4 → 0 [MEASURED]; - the band table repairs: [300,500) reads 0.480 vs measured 0.520 (D4 said 0.423); [500,710) reads **0.351 vs measured 0.357** (D4 said 0.233) [MEASURED]; - held-out deep PLN improves **+3.6 nats with sharper bands** (model sd 0.0834 vs the record's 0.177) [MEASURED]. **What remains is a constant-rate exponential δ″ ≈ 0.19–0.20 in θ/m̄, now nearly depth-stable, and underived.** That is the sharpened open object: not "δ = 0.277 with a flattening", but a flat ~0.19 that no singular-series layer, no derived-null drift, and no finite-size term supplies. **(b) x = 37: the HL layer is unexceptional, cleanly, and the anomaly localises to G2 being LARGE while h and h2 are exactly normal.** W1_37(528) = 2.571 sits at 1.044× its exact Gallagher mean, 58th percentile of its own neighbourhood, mid-pack in a level range [0%, 93%] where the comb-loudest levels (61, 79 at 1.84×) and comb-mutest (67 at 0.40×, W1 = 1 exactly) are all non-anomalous. The comb-corrected threshold keeps the full spike (c2″(37) = 0.5732, 0 of 10 later terms above). Leave-one-out log-midpoint residuals: **G2 at +2.39 series-noise units, h at +0.14, h2 at −0.10**. With the anchored/mirror negative (`attack-anchored-01.md` §5), this closes the **third** unexceptional mechanism family at 37. The spike stays real and stays unexplained. ## 1. Custody and calibration Nothing re-sieved. The per-fold field (102 folds, p ∈ [101, 709]) is parsed from `attack-mp-derive-01.js`'s embedded Stage-6 table; W1 is recomputed exactly from (p, θ) and matches the printed column at worst 0.0050 (the printing figure w1hl-01 recorded); the 65 deep Sλ are upgraded to 5-figure precision from the window-5 B2 embed via the record's own Sλ = 1.1211·λ₁₁ identity; the ladders come from `a144311-full-ladder.js` and `external-ladders-01.js` embeds with verbatim spot gates. Calibration gates, all passing before any candidate is scored: the record's (k, δ) = (1.9468, 0.2771) reproduce as (1.9485, 0.2773), deep MLE 0.2526 vs 0.252, the Sλ column carries the law's fixed exponent (slope −1.0866 vs c = 1.0818 [foldL-06 §3.1]), and the PLN engine lands 0.6 nat from the record's −88.6 (grid-independent input rounding; cancels in every model difference). ## 2. The finding: the interior 6-tuple comb is the residual sub-field `mp-derivation.md` §1 modeled W1 as "endpoints exact, interior smooth". The next singular-series layer drops the second half. For a gap of value θ at fold p, condition each interior slot j ∈ {6, 12, …, θ−6} on the endpoint quadruple {0, 2, θ, θ+2} being alive: per prime q ∈ [5, p), with the offset uniform, **[PROVEN, elementary]** > P(j alive | endpoints alive) / P(generic slot alive) > = [(q − ν₆)/(q − ν₄)] · [q/(q − 2)], > ν₆ = |{0, −2, −j, −j−2, −θ, −θ−2} mod q|, ν₄ the endpoint count, and the first-order correction to the exactly-θ gap count is exp(−(6/m̄)·D), D = Σ_j (∏_q r_q(j) − 1). The per-prime **mean over a full residue system of j is exactly 1** (verified in-code at q ∈ {7,13,31,97} × θ ∈ {204,660,1260}, and it is one line: E_j[P(j alive | end)] = (q−2)/q regardless of the endpoint class) — so Gallagher mean-one kills the effect in the infinite limit, and **the entire correction is the finite interior range 6…θ−6 breaking equidistribution**: for q > θ/6 the interior slots hit θ/6 − 1 distinct residues that deterministically exclude the two strongest enhancing classes (j ≡ 0 and j ≡ θ). It is the same mechanism shape as the prime-indexed mean shift of `w1-singular-series.md` §2, one layer up. This zero-parameter arithmetic field, evaluated exactly (about 3·10⁶ residue-set counts, 0.5 s), correlates 0.907 with the measured D4 train residuals and halves the train roughness. It is not comb-collinear (corr(ℓ, ln W1) = −0.013 over the 102 folds), so it is genuinely the *next* term, not a W1 echo. Among the four flags it most helps 409 (ℓ = 0.4554 vs field mean 0.3603); combined with the null credit all four drop below 3. **Calibration note on "predicts the flattening".** The interior comb's band-OLS tilt difference is φ_pred = 0.0241 vs the measured 0.0247 — on the nose as a band summary. Under the record's own estimators (train OLS + deep exposure-weighted MLE) crediting (i-b) alone moves φ 0.0247 → 0.0191, and crediting (ii)+(i-b) moves it to 0.0085: the deep MLE weights the shallow end of the deep band (Sλ falls from 111 to 0.06 across it), so band-OLS summaries and MLE refits are deliberately both reported, and the refit is the honest test. ## 3. The two structural candidates, priced **(ii) The derived null.** Its depth structure is two opposing drifts, both computed exactly per fold and custody-licensed (the normalisation X_null = (Y/m̄)(2/p)·r_null reproduces import-thinning §1.4's zero-parameter window predictions: total 5481.9 exact, decades to 0.1–0.2%): the exponent drift (c_law − c_null(p))·z is *anti-decay* (+0.0643/z — c_null falls 1.0623 → 1.0298 across the field, always below the fitted 1.0818), while the derived amplitude 3/m̄ — the prefactor THN §0 records the law freezes into a constant A — is decay (−0.1090/z train, fading to −0.0452 deep). Net: −0.0600/z, right sign, 22% of δ, and a right-signed φ. Crediting it exactly leaves δ′ = 0.2172/0.2031. The composed CRT second-order bracket (exponents 1.0302…1.1096 at u = 0.25…0.9, THN §4.3, cited not recomputed) cannot reach the remaining ~0.19 either. **(i-a) Published second moments.** No channel, structurally: the field's comb argument is deterministic per fold and carried exactly, so integer-ensemble variance results bear on spread the model already owns. The two honest numbers: the truncation drift of E_v[W1_P] across the whole field is 0.67% in ln against a needed −2.67 nats, and the exact second-moment spread (sd ≈ 0.465 lognormal-read) sits near the measured sd(ln W1) = 0.515. Prior-art position of the second-moment layer: see §5. **(iii) Finite-size/edge.** Derived exactly and refuted by orders of magnitude: window-edge ≤ 2.5·10⁻⁸ pooled (≤ 2.8·10⁻⁴ at the worst single window), W1-cutoff tail (primes q ≥ p dividing θ(θ²−4): q = p always, factor (p−3)/(p−4), plus any prime factor > p of (θ±2)/2, factored exactly) mean 4.3·10⁻³, worst 2.0·10⁻², tilt −0.0024/z. ## 4. x = 37, the computed layer Fold comb θ_37 = 72 = 2³·3²: no folded divisor ≥ 5 at all — the whole comb is W1(72) = 2·(4/3) = 2.667 from the θ−2 side (70 = 2·5·7), mid-band. Tile comb at the record gap: W1_37(528) = (9/7)·2 = 2.571 (11 | 528; 5 | 530), ratio 1.044 to the exact integer-indexed mean E_v[W1_37] = 2.462, 58th percentile among v ≡ 0 (6) in [0.7·G2, 1.3·G2]. The level table (embed, 6a) shows comb ratio and percentile are uncorrelated with anomaly status across x = 11…79. c2″ = G2/(m·(lnD + ln W1(G2))) moves the spike by ~3.6% against its ~19% size and 37 keeps 0 of 10 later terms above it. The three-series LOO residuals give the sharpest one-line statement yet of the instrumental reading: G2 +2.39 noise units, h +0.14, h2 −0.10 — **the anomaly is G2 large, not h2 small**, and the h2-side is now quantified as exactly normal. ## 5. Prior art (searched per `SEARCH-CONVENTIONS.md`), and the Wolf sweep - **Montgomery–Soundararajan, "Primes in short intervals", arXiv math/0409258 — read at page images (pp. 1–4), hypotheses in full.** Their Theorem 1 (eq. 15: V_k(q;h) = μ_k V₂(q;h)^{k/2} + O_k(h^{k/2−1/(7k)} (q/φ(q))^{2^k+k/2}), k any positive integer) and Theorem 2 (R_k(h) = Σ_{distinct d_i ≤ h} 𝔖₀(D) = μ_k(−h log h + Ah)^{k/2} + O_k(h^{k/2−1/(7k)+ε}), A = 2 − C₀ − log 2π; R₂(h) = −h log h + Ah + O(h^{1/2+ε}), their (16)) are **unconditional** statements about moments of the deleted singular series averaged over integer tuples in [1, h]. Their conclusion about ψ(x+H)−ψ(x) being normal with variance ~H log(N/H) is **conditional on a strong uniform form of the Hardy–Littlewood conjecture (their (1))** — stated so the import is not mistaken for coverage. Nothing in either theorem indexes an average by fold depth or by shifted primes; the no-channel verdict of §3 is our structural argument, not their theorem. - **The Wolf sweep — the named unswept hiding place is now swept, and the W1 comb is NOT there.** Two full texts read as page images: (i) M. Wolf, *Some Remarks on the Distribution of twin Primes*, arXiv:math/0105211 (2001; the write-up of his 1990s computations to 2⁴⁴ ≈ 1.76·10¹³, and the direct response to Kelly–Pilling). His m(d, N) — counts of consecutive twins at arithmetic separation d — is plotted with the comb's oscillations plainly visible (his Fig. 1), he names the twin separation champions d = 30, then d = 210, promises "the forthcoming paper" on them, and then derives formulas ONLY for the prime-count separation μ(s, N) ~ A(N)e^{−B(N)s} (A ≈ c₂²N/ln³N, B ≈ c₂/ln N), noting that the change of measuring sticks "removes oscillations from Fig.1 and leaves pure exponential decrease". **The comb's shadow is in his figure; no divisor-modulated formula for twin separations appears anywhere in the paper.** His s_max(N) ~ (1/c₂)ln²N and the remark that arithmetic maximal twin gaps grow ~ln³N are the Kelly–Pilling-adjacent objects, comb-free. (ii) M. Wolf, *On the Twin and Cousin Primes*, IFTUWr 909/96 (August 1996, retrieved from his UKSW publication page as PDF): entirely π₂(x) vs π₄(x) — the fractal sign-change function W(x) = π₂ − π₄, random-walk analogies, generalized Brun constants. No separation statistics at all. (iii) Not opened: IFTUWr 894/95 (*Some conjectures on the gaps between consecutive primes*, PostScript only) — that paper is about PRIME gaps, where Wolf's divisor factor is the standard pair comb 𝔖₂(d) (the corpus's W0 shape), already indexed as standard; no twin-twin claim rests on it. A "twin champions" follow-up never appears on his publication list (checked 2026-08-21); the nearest is Odlyzko–Rubinstein– Wolf, *Jumping Champions* (Exp. Math. 8, 1999), which is primes, not twins. **The `w1-singular-series.md` §3 NOT-FOUND on the explicit W1 comb therefore survives its last named hiding place.** - **The interior-comb object itself (i-b).** The inclusion–exclusion of interior points with higher singular series is the standard apparatus of consecutive-gap heuristics (it is how gap counts between consecutive primes are corrected; Brent's and Odlyzko-style computations); our §2 per-prime ratio is a two-line instance of it. What was not found in print: this correction used as a per-fold field on twin-comb fold depth, or any statement of its finite-range mean-one breaking. As with W1: the algebra is standard, the application is ours. (One-session sweep; zbMATH/MathSciNet not consulted.) ## 6. Proposed consequences, for adjudication rather than applied - `mp-derivation.md` §5's "a real residual sub-field (sd ≈ 0.18, a far-tail flattening, fold 631, four flagged folds) is deterministic-looking and NOT derived" can now carry: *derived — the sub-field is the second singular-series layer (interior 6-tuple comb, zero parameters, corr 0.907, sd 0.18 → 0.083) plus the thinning null's own depth drift; all four flags fall below 3 and the flattening collapses to 0.0085 (`attack-delta37-01.md`). The remnant is a nearly depth-stable exponential δ″ ≈ 0.19–0.20, still underived.* - `w1-singular-series.md` §5's Wolf line can close: *swept 2026-08-21 at page images (math/0105211 + IFTUWr 909/96): oscillations observed, champions named, no comb written — the NOT-FOUND stands.* - The δ hunt's live target sharpens: the ~0.19 flat remnant, with the first-order interior layer now exhausted; the natural next objects are the interior-PAIR (second-order in 6/m̄) terms and the measured steeper-than-geometric hazard structure of THN §1.5 — i.e. CRT dependence beyond every first-order layer priced here. - If adjudicated in, the credited model is a strictly sharper prereg RIVAL line (sd 0.083 vs 0.177) under the mp-window prereg's own consequence rule (formula bands are rivals, never replacements). ## 7. What this does not show δ″ ≈ 0.19–0.20 remains fitted, not derived; nothing here touches H″, the exponent, or any live grade. The interior correction is first order in the level density (interior-PAIR terms not computed) and prices only the dominant qualifying value θ_c (the 4p and 6p channels carry their own e^{−z}-suppressed interiors). The 0.907 correlation is a train-band statement (37 folds); deep folds enter only through exposure-limited MLE/PLN scores, and the blind-window rows were not rescored. The credited k′ values absorb each tilt's constant level and are not comparable to the record's k. The x = 37 negative closes the HL mechanism family for the *level instruments* (c2′, G2/h, h2/G2); it does not explain the spike. ## 8. Reproduction ``` node research/attack-delta37-01.js # 0.6 s, no sieving node research/qc/embed.js --check research/attack-delta37-01.js ``` --- *This document states current understanding. Superseded claims, retired numbers and the reasons they changed are in [../CHANGELOG.md](../CHANGELOG.md), indexed by document.*