#!/usr/bin/env node // ============================================================================ // IMPORT-SUEN 01 — THE CORRELATION-INEQUALITY TRANSFER AT THE ANCHOR // (2026-08-19, foreign-import attack 2 of 5: Suen / Janson / Lovász Local // Lemma against the anchored dependence term δ) // ============================================================================ // THE HOLE THIS FIRES AT. `research/history/staging/verify-cofactor-convolution.md` // §9 says verbatim: "No proof that δ = O(Σ_{q>x} q^{−2}): the averaging over // mixed patterns is uncontrolled and the @19 sign flip is unexplained." // δ is defined in that file's §7 by S(0) = N̄·p₀·q₀·(1 + δ), and the report // measures |δ| = 0.29%..2.02% against a "forced CRT scale" // F = ∏(1−1/q)²/∏(1−2/q) − 1 = Σ_q 1/(q(q−2)) + … = 0.88%..2.11%. // // OBJECTS, from `paper/anchored-note.md` §1, imported by definition and not by // code. Level x; tile W = x#; the Natal@5 comb // N = { r ∈ [0,W) : r ≡ 11 or 17 (mod 30), r mod p ∉ {0, p−2} for 7 ≤ p ≤ x }, // |N| = N̄ = 2·∏_{7≤p≤x}(p−2); scour depth y = largest prime ≤ √W; scour primes // are the primes q with x < q ≤ y. At the anchor t = 0 the prime q strikes r iff // q | r or q | r+2. Write ω_s(r) = #{scour q : q | r}. Then // L0 = #{r ∈ N : ω_s(r) = 0}, R0 = #{r ∈ N : ω_s(r+2) = 0}, // S0 = #{r ∈ N : ω_s(r) = ω_s(r+2) = 0} = S(0), // p₀ = L0/N̄, q₀ = R0/N̄, 1 + δ = S0·N̄ / (L0·R0). // A natal r has no prime factor ≤ x, and a composite natal r < W must have a // prime factor ≤ y (two factors above y would exceed y'² > W). So // ω_s(r) = 0 ⟺ r is prime AND r > y, // the second clause because a natal prime r ≤ y IS a scour prime and q = r // strikes r. That clause is the whole difference between 47 and the published // S(11) = 45; dropping it was this file's first draft and it is recorded here // because the wrong version looks right. Likewise ω_s(r+2) = 0 ⟺ r+2 prime and // r+2 > y. So L0, R0, S0 are prime counts on the comb and this file computes // them by a segmented odd-only sieve — no import from // research/natal-cap-35-x-multiplicity.js, whose figures this file re-derives. // // THE MAPPING (stage 1). Ground set: for each scour prime q the residue // X_q = r mod q. Bad events A_q^L = {X_q = 0}, A_q^R = {X_q = −2}. For r uniform // modulo ∏_q q the X_q are EXACTLY independent by CRT, and the natal condition // lives mod x#, which is coprime to every scour prime, so the natal comb is // exactly independent of the whole scour vector. On that product measure ν: // • two bad events are dependent iff they share a PRIME, never a slot; // • the dependency graph is a perfect matching on 2K vertices, max degree 1; // • every edge is a mutual exclusion (q | r and q | r+2 would force q | 2), so // Janson's Δ = Σ_{i∼j} P(B_i ∧ B_j) = 0 EXACTLY. // Hence on ν every member of the family (Janson, Suen, lopsided LLL, Harris) is // exactly tight and adds nothing: P_ν(no strike) = ∏(1−2/q) by factorisation. // PART B measures that vacuity rather than asserting it. // // AND JANSON DOES NOT MERELY GO VACUOUS, IT DOES NOT APPLY. Janson's inequality // (Alon-Spencer Thm 8.1.1) is stated for events B_i = {A_i ⊆ R} on a random // subset R with independent inclusions, i.e. for INCREASING events, and it // asserts ∏P(B̄_i) ≤ P(∧B̄_i) as its lower half. Here P(∧B̄) = ∏(1−2/q) is // SMALLER than ∏P(B̄_i) = ∏(1−1/q)², by exactly the factor 1/(1+F), so the // lower half is false for our events. The reason is structural and is worth one // line: two nonempty up-sets in a finite product lattice always share the top // element, so two DISJOINT nonempty increasing events cannot exist. A_q^L and // A_q^R are disjoint and nonempty. No choice of ground set puts the two strike // orientations into Janson form. The mutual exclusion that makes the problem's // arithmetic pleasant is exactly what puts it outside the theorem. // // PART E therefore runs the one member of the family that does NOT need a // product space — Suen's inequality on a superdependency digraph, Janson (1998), // "New versions of Suen's correlation inequality", RSA 13, 467-483, Theorems 2, // 3 and 8 — against the anchored measure, where the only valid dependency graph // is the complete one, and prints how vacuous each bound is. // // THE TRANSFER (stage 2). The anchored measure is r uniform on N ⊂ [0,W), and // its scour-residue vector is NOT the product measure, because W = x# is smaller // than ∏_{xx} // (1 − 1/(p−1)²), which is 1/(1+F) up to the q > y tail. So the honest // prediction for δ is −F/(1+F): the forced scale with a MINUS sign and // no free constant. Measured δ is smaller than that at four of five // levels, so ρ > 0 is expected at every level. // // THE WALL (stage 4). PART D puts the covering question — "some slot in a window // of length H survives all q ≤ x" — in local-lemma coordinates. The lopsided LLL // hypothesis is P(A_q | ∩_{q'∈S} Ā_{q'}) ≤ x_q ∏_{q'∼q}(1−x_{q'}) quantified over // ALL subsets S of the non-neighbours of q. Verifying it for a window of length H // needs equidistribution of that window modulo ∏_{q'∈S} q'. PART D computes the // three thresholds this produces. // ============================================================================ 'use strict'; const f = (v, d = 4) => Number.isFinite(v) ? v.toFixed(d) : String(v); const pct = (v, d = 4) => Number.isFinite(v) ? (100 * v).toFixed(d) + '%' : String(v); function primesUpTo(n) { const s = new Uint8Array(n + 1), o = []; for (let i = 2; i <= n; i++) { if (!s[i]) { o.push(i); for (let j = i * i; j <= n; j += i) s[j] = 1; } } return o; } // --------------------------------------------------------------------------- // PART A — exact anchored counts, by segmented odd-only sieve // --------------------------------------------------------------------------- // For level x we need, for every natal r ∈ [0, W): is r prime, is r+2 prime. // Odd-only segmentation; the natal mask is sieved in the same segment rather // than tested by division, so the cost is O(W loglog W) with a small constant. function anchoredCounts(x) { const base = primesUpTo(x).filter(p => p >= 7); // the wheel primes 7..x let W = 30; for (const p of base) W *= p; const Nbar = 2 * base.reduce((a, p) => a * (p - 2), 1); const root = Math.floor(Math.sqrt(W + 2)); const bps = primesUpTo(root).filter(p => p > 2); // odd base primes for the sieve const y = bps[bps.length - 1]; // largest prime ≤ √W const scour = primesUpTo(y).filter(q => q > x); const SEGO = 1 << 22; // odds per segment const comp = new Uint8Array(SEGO + 2); // primality of the segment's odds (+2 slack) const nat = new Uint8Array(SEGO + 1); // 0 = natal, 1 = not let Ncheck = 0, L0 = 0, R0 = 0, S0 = 0; const nseg = Math.ceil(W / (2 * SEGO)); for (let s = 0; s < nseg; s++) { const lo = 1 + 2 * s * SEGO; // first odd of the segment const hiOdd = lo + 2 * SEGO; // one past, as an odd comp.fill(0); // 1 and any odd < 2 handled below by the explicit r ≥ 7 natal filter. for (const p of bps) { let start = p * p; if (start < lo) { start = Math.ceil(lo / p) * p; if ((start & 1) === 0) start += p; } if (start >= hiOdd + 2) continue; for (let i = (start - lo) / 2; i <= SEGO + 1; i += p) comp[i] = 1; } // natal mask: start "not natal" everywhere, clear the 11 and 17 mod 30 odds, // then re-mark the two forbidden classes of every wheel prime. nat.fill(1); for (const c of [11, 17]) { let v = lo + ((c - (lo % 30)) % 30 + 30) % 30; // first ≡ c (mod 30) at or after lo for (let i = (v - lo) / 2; i < SEGO; i += 15) nat[i] = 0; } for (const p of base) { const iv = (p + 1) / 2; // inverse of 2 mod p for (const cls of [0, p - 2]) { // r = lo + 2i ≡ cls (mod p) ⟺ i ≡ (cls − lo)·iv (mod p) let i0 = ((((cls - lo) % p) + p) % p) * iv % p; for (let i = i0; i < SEGO; i += p) nat[i] = 1; } } const lim = Math.min(SEGO, Math.ceil((W - lo) / 2)); for (let i = 0; i < lim; i++) { if (nat[i]) continue; Ncheck++; const r = lo + 2 * i; const a = !comp[i] && r > y, b = !comp[i + 1] && r + 2 > y; // survives the scour if (a) L0++; if (b) R0++; if (a && b) S0++; } } return { x, W, Nbar, Ncheck, y, K: scour.length, scour, L0, R0, S0 }; } // --------------------------------------------------------------------------- // PART B/C helpers — the product measure ν and the split // --------------------------------------------------------------------------- function nuStats(scour) { let lnP1 = 0, lnP2 = 0, lnF = 0, s_q2 = 0, s_qm1 = 0, s_qqm2 = 0, mu = 0, s4q2 = 0; for (const q of scour) { lnP1 += Math.log(1 - 1 / q); // ∏(1−1/q) lnP2 += Math.log(1 - 2 / q); // ∏(1−2/q) lnF += Math.log(1 + 1 / (q * (q - 2))); // ∏(1−1/q)²/∏(1−2/q) s_q2 += 1 / (q * q); s_qm1 += 1 / ((q - 1) * (q - 1)); s_qqm2 += 1 / (q * (q - 2)); mu += 2 / q; s4q2 += 4 / (q * q); } return { P1: Math.exp(lnP1), P2: Math.exp(lnP2), F: Math.expm1(lnF), // the forced scale, exactly s_q2, s_qm1, s_qqm2, mu, DeltaMatch: 0, // true dependency graph: mutual exclusion DeltaComplete: mu * mu - s4q2, // one-event-per-prime reading }; } // tail ∏_{p>y}(1 − 1/(p−1)²) as 1 − Σ_{p>y} 1/(p−1)², estimated by ∫ dt/(t ln t · t) function tailProd(y) { return 1 - 1 / (y * Math.log(y)); } // --------------------------------------------------------------------------- // PART D — the wall in local-lemma coordinates // --------------------------------------------------------------------------- // (i) PAIRWISE HORIZON. Draw the dependency graph the way one draws it: join q, // q' iff the window of length H fails to equidistribute mod q·q', i.e. iff // q·q' > H. With H ≥ x² no pair of primes ≤ x is joined and the graph is // EMPTY, so the LLL would certify a survivor at H = x². That is exactly the // p² rule. The hypothesis, however, quantifies over all SETS S. // (ii) SET-WISE / COMPLETE GRAPH. If no set beyond a pair can be verified, the // graph must be taken complete and the asymmetric LLL condition // 2/q ≤ x_q ∏_{q'≠q}(1−x_{q'}) with x_q = c/q becomes 2 ≤ c·∏_{p≤x}(1−c/p), // which is the Mertens wall. Reported below as the largest feasible x. // (iii) THE ADMISSIBLE REPAIR. Verify the hypothesis only for sets S with // ∏_{q∈S} q ≤ H, i.e. |S| ≤ m ≈ ln H/ln x. That is Bonferroni truncation at // depth m, i.e. Brun's pure sieve, and the exponent it certifies is computed // below and compared with `research/sift-limit-attack.md` §7's β_pure. function lllWall() { const rows = []; for (const x of [5, 7, 11, 13, 17, 19, 23, 29, 101, 199, 227, 229, 439, 1009, 1000003]) { const ps = primesUpTo(x).filter(p => p >= 5); // (ii) complete-graph asymmetric LLL with weights c/p: max_c c·∏(1−c/p) vs 2 let best = 0, bestc = 0; for (let c = 0.01; c <= 8.0001; c += 0.01) { let lg = 0, ok = true, worst = -Infinity; for (const p of ps) { if (c / p >= 1) { ok = false; break; } lg += Math.log(1 - c / p); worst = Math.max(worst, Math.log(1 - c / p)); } if (!ok) continue; // condition for every p: 2 <= c * prod_{p' != p}(1 - c/p'); the binding p is // the one whose own factor is largest, i.e. the largest prime. const v = c * Math.exp(lg - worst); if (v > best) { best = v; bestc = c; } } const mert = ps.reduce((a, p) => a + 2 / p, 0); // (iii) Bonferroni depth: smallest even m with e_{m+1} < ∏(1−2/p), e_j the // elementary symmetric functions of {2/p}. Computed exactly, not by Stirling. let e = [1]; for (const p of ps) { const t = 2 / p, ne = e.slice(); ne.push(0); for (let j = e.length; j >= 1; j--) ne[j] = e[j - 1] * t + (e[j] || 0); e = ne; } const prod2 = ps.reduce((a, p) => a * (1 - 2 / p), 1); while (e.length < 4) e.push(0); let m0 = -1; for (let m = 0; m + 1 < e.length; m += 2) { if (e[m + 1] < prod2) { m0 = m; break; } } // exponent: the depth-m error term is Σ_{j≤m} 2^j C(π,j) ≈ (2π(x))^m/m!, so // ln H ≥ ln(that / main), and θ = ln H / ln x. let theta = NaN; if (m0 >= 0) { const pi = ps.length; let lnErr = 0; for (let j = 1; j <= m0; j++) lnErr = Math.max(lnErr, j * Math.log(2 * pi / j) + j); theta = (lnErr - Math.log(Math.max(prod2 - e[m0 + 1], 1e-300))) / Math.log(x); } rows.push({ x, mert, bestc, best, feasible: best >= 2, m0, theta, pairwise: 2 }); } return rows; } // --------------------------------------------------------------------------- // PART E — Suen's inequality, actually applied, on the anchored measure // --------------------------------------------------------------------------- // Janson (1998) Thm 2: P(S=0) ≤ exp(−μ + Δ·e^{2δ}) // Thm 3: P(S=0) ≤ exp(−min(μ²/8Δ, μ/6δ, μ/2)) // Thm 8: P(S=0) ≥ (1 − Δ₀*·e^{Δ*})·∏(1−p_k) // with one event per scour prime, I_q = 1{q | r(r+2)}, p_q = 2/q; μ = Σp_q; // Δ = Σ_{{i,j}: i∼j} E(I_iI_j) over UNORDERED pairs (Janson's Remark 4 warns of // the factor-2 convention); δ = max_i Σ_{j∼i} p_j. // // THE GRAPH MUST BE COMPLETE. Janson's Remark 3 requires a dependency graph in // the STRONG sense and gives an explicit counterexample for the pairwise notion. // On the anchored measure the residues (r mod q) of r ∈ [0,W) are jointly // dependent for every subset of scour primes with ∏q > W, so no proper subgraph // is admissible and the complete graph is the only valid choice. Then // ∏_{k∼{i,j}}(1−p_k)^{−1} = 1/∏_k(1−p_k) — both endpoints are neighbours of // {i,j}, so the pair's neighbourhood is every other prime — which is what makes // Thm 8's Δ* explode. E(I_iI_j) is taken at its ν value p_ip_j; the true // anchored pair correlations differ by percents and the vacuity margin below is // tens of orders of magnitude, so the substitution changes nothing. function suen(scour, P2) { let mu = 0, s2 = 0; for (const q of scour) { mu += 2 / q; s2 += 4 / (q * q); } const Delta = (mu * mu - s2) / 2; // unordered pairs const dmax = mu - 2 / scour[scour.length - 1]; // max_i sum_{j~i} p_j // Δ* and Δ₀* with the complete graph: weight 1/P2 on each pair, so the inner // sum is the unordered-pair Δ computed above: Σ_{i `@${a.x}: ${f(B[i].DeltaComplete / B[i].F, 1)}×`).join(' ')); console.log('\n=== PART C — the exact split (1+δ)(1+F) = 1+ρ, and the HL prediction ==='); console.log('x | δ measured | F forced | δ_HL=−F/(1+F) | ρ = β/(π_Lπ_R)−1 | ρ/F | ρ−F | identity'); const C = []; for (let i = 0; i < A.length; i++) { const a = A[i], n = B[i]; const delta = (a.S0 * a.Nbar) / (a.L0 * a.R0) - 1; const rho = (1 + delta) * (1 + n.F) - 1; const dHL = 1 / (1 + n.F) - 1; const dHLtail = tailProd(a.y) / (1 + n.F) - 1; C.push({ delta, rho, dHL, dHLtail, F: n.F }); console.log([ String(a.x).padStart(2), pct(delta, 4).padStart(10), pct(n.F, 4).padStart(9), pct(dHL, 4).padStart(13), pct(rho, 4).padStart(16), f(rho / n.F, 3).padStart(5), pct(rho - n.F, 4).padStart(9), ((rho > n.F) === (delta > 0) ? 'ok' : 'ARITHMETIC FAIL') ].join(' | ')); } console.log('cross-check against the three published ratios (β, π_L, π_R) — computed here, not imported:'); console.log('x | β = S0/(N̄∏(1−2/q)) | π_L | π_R | π_L·π_R | β/(π_Lπ_R)−1 = ρ'); for (let i = 0; i < A.length; i++) { const a = A[i], n = B[i]; const beta = a.S0 / (a.Nbar * n.P2); const pL = (a.L0 / a.Nbar) / n.P1, pR = (a.R0 / a.Nbar) / n.P1; console.log([ String(a.x).padStart(2), f(beta, 5).padStart(19), f(pL, 5).padStart(8), f(pR, 5).padStart(8), f(pL * pR, 5).padStart(8), pct(beta / (pL * pR) - 1, 4).padStart(16) ].join(' | ')); } console.log('HL prediction with the q > y tail restored, δ_HL = ∏_{p>x}(1−1/(p−1)²) − 1:'); console.log(' ' + A.map((a, i) => `@${a.x}: ${pct(C[i].dHLtail, 4)}`).join(' ')); console.log('measured δ minus HL-predicted δ (this difference IS the HL relative error at W = x#):'); console.log(' ' + A.map((a, i) => `@${a.x}: ${pct(C[i].delta - C[i].dHLtail, 4)}`).join(' ')); console.log('\n=== PART E — Suen (Janson 1998) applied, complete graph, anchored measure ==='); console.log('x | mu | Δ (unord) | δ=max_i Σ_{j~i}p_j | Δ* | Thm2 exponent | Thm3 bound | Thm8 bracket | truth S0/N̄'); for (let i = 0; i < A.length; i++) { const a = A[i], n = B[i]; const S = suen(a.scour, n.P2); console.log([ String(a.x).padStart(2), f(S.mu, 5).padStart(7), f(S.Delta, 4).padStart(9), f(S.dmax, 5).padStart(18), f(S.DeltaStar, 3).padStart(9), ('+' + f(S.t2, 1)).padStart(13), f(Math.exp(S.t3), 5).padStart(10), S.t8.toExponential(2).padStart(12), f(a.S0 / a.Nbar, 6).padStart(11) ].join(' | ')); } console.log('Thm 2 exponent is positive at every level, so its bound exceeds 1 and says nothing.'); console.log('Thm 3 is an UPPER bound on the survival probability, the wrong direction for'); console.log('Assumption A, and it is loose by the ratio printed next:'); console.log(' ' + A.map((a, i) => `@${a.x}: ${f(Math.exp(suen(a.scour, B[i].P2).t3) / (a.S0 / a.Nbar), 2)}×`).join(' ')); console.log('Thm 8 is the lower bound the programme needs. Its bracket 1 − Δ₀*e^{Δ*} is'); console.log('positive only at @11 and negative from @13 on, by the magnitudes printed next'); console.log('(a negative bracket means the theorem asserts P(S=0) >= a negative number):'); console.log(' ' + A.map((a, i) => { const S = suen(a.scour, B[i].P2); return `@${a.x}: 10^${f(Math.log10(Math.abs(S.t8)), 1)}`; }).join(' ')); console.log('\n=== PART D — the wall in local-lemma coordinates ==='); console.log('(i) pairwise horizon: graph empty iff q·q\' ≤ H for all q,q\' ≤ x, i.e. H ≥ x². θ = 2.'); console.log('(ii) complete graph (set-wise hypothesis unverifiable): need max_c c·∏_{5≤p≤x}(1−c/p) ≥ 2.'); console.log('(iii) admissible repair: |S| ≤ m with x^m ≤ H, i.e. Bonferroni depth m, i.e. Brun pure sieve.'); console.log('x | Σ_{5≤p≤x}2/p | best c | max c∏(1−c/p) | (ii) feasible | m₀ (depth) | θ_pure = lnH/lnx'); for (const r of lllWall()) { console.log([ String(r.x).padStart(7), f(r.mert, 4).padStart(12), f(r.bestc, 2).padStart(6), f(r.best, 5).padStart(13), (r.feasible ? 'yes' : 'NO').padStart(13), String(r.m0).padStart(10), f(r.theta, 3).padStart(16) ].join(' | ')); } console.log('θ targets for comparison: pairwise horizon 2 (= the p² rule); best published'); console.log('dimension-2 sifting limit β₂ = 4.26645028414864191641 (research/SEARCH-CONVENTIONS.md §4).'); // ============================================================================ // OUTPUT — EMBEDDED, do not hand-edit. Regenerate: // node research/qc/embed.js research/import-suen-01-transfer.js -- 11,13,17,19,23,29,31 // invocation: node research/import-suen-01-transfer.js 11,13,17,19,23,29,31 // code-sha256: ffa1cedada93ea212f2027a477f06d37648482d530b308043e8581d310dc7ad6 // out-sha256: 9f96369166ece9a4acf84078b901ea9559f377e57c7d3ec8a822492b42818166 // body-lines: 90 // forced: 2026-09-25, 16 of 289 figures in the replaced block not reproduced (first: 0.518, 1.31e-1, 1.826, -1.03e+1) // streams: stdout // node: v22.23.3 // embedded: 2026-09-25 // elapsed: 589.9 s // ============================================================================ // IMPORT-SUEN 01 — correlation inequalities against the anchored dependence δ // levels: 11, 13, 17, 19, 23, 29, 31 // // === PART A — exact anchored counts (independent segmented sieve) === // x | W | N̄ | y | K | L0 | R0 | S0 | N̄ ok // 11 | 2310 | 90 | 47 | 10 | 62 | 64 | 45 | PASS // 13 | 30030 | 990 | 173 | 34 | 558 | 547 | 307 | PASS // 17 | 510510 | 14850 | 709 | 120 | 6813 | 6775 | 3099 | PASS // 19 | 9699690 | 252450 | 3109 | 435 | 98340 | 98245 | 38380 | PASS // 23 | 223092870 | 5301450 | 14929 | 1739 | 1784710 | 1783974 | 597475 | PASS // 29 | 6469693230 | 143139150 | 80429 | 7863 | 42074348 | 42072589 | 12307838 | PASS // 31 | 200560490130 | 4151035350 | 447829 | 37534 | 1087136413 | 1087113427 | 283449187 | PASS // // === PART B — the product measure ν: dependency structure and Janson Δ === // true dependency graph: perfect matching on 2K vertices, max degree 1, // every edge a mutual exclusion, so Janson Δ_match = 0 exactly at every level. // x | K | mu=Σ2/q | Δ_match | Δ_complete | F (forced scale) | Σ1/(q(q−2)) | Σ1/q² | Σ1/(q−1)² // 11 | 10 | 0.78909 | 0.0 | 0.5483 | 2.1086% | 2.0908% | 1.8586% | 2.0826% // 13 | 34 | 1.14676 | 0.0 | 1.2525 | 1.7084% | 1.6956% | 1.5632% | 1.6922% // 17 | 120 | 1.49379 | 0.0 | 2.1798 | 1.3856% | 1.3771% | 1.2902% | 1.3752% // 19 | 435 | 1.78566 | 0.0 | 3.1475 | 1.0878% | 1.0824% | 1.0281% | 1.0814% // 23 | 1739 | 2.05259 | 0.0 | 4.1795 | 0.8819% | 0.8783% | 0.8420% | 0.8778% // 29 | 7863 | 2.30539 | 0.0 | 5.2859 | 0.7537% | 0.7511% | 0.7236% | 0.7507% // 31 | 37534 | 2.52361 | 0.0 | 6.3438 | 0.6419% | 0.6399% | 0.6196% | 0.6397% // ratio Δ_complete / F, i.e. how far Janson's error term sits from the forced scale: // @11: 26.0× @13: 73.3× @17: 157.3× @19: 289.3× @23: 473.9× @29: 701.3× @31: 988.3× // // === PART C — the exact split (1+δ)(1+F) = 1+ρ, and the HL prediction === // x | δ measured | F forced | δ_HL=−F/(1+F) | ρ = β/(π_Lπ_R)−1 | ρ/F | ρ−F | identity // 11 | 2.0665% | 2.1086% | -2.0651% | 4.2187% | 2.001 | 2.1101% | ok // 13 | -0.4246% | 1.7084% | -1.6797% | 1.2765% | 0.747 | -0.4319% | ok // 17 | -0.2988% | 1.3856% | -1.3667% | 1.0827% | 0.781 | -0.3030% | ok // 19 | 0.2859% | 1.0878% | -1.0761% | 1.3767% | 1.266 | 0.2890% | ok // 23 | -0.5149% | 0.8819% | -0.8742% | 0.3625% | 0.411 | -0.5194% | ok // 29 | -0.4770% | 0.7537% | -0.7481% | 0.2732% | 0.362 | -0.4806% | ok // 31 | -0.4428% | 0.6419% | -0.6378% | 0.1963% | 0.306 | -0.4456% | ok // cross-check against the three published ratios (β, π_L, π_R) — computed here, not imported: // x | β = S0/(N̄∏(1−2/q)) | π_L | π_R | π_L·π_R | β/(π_Lπ_R)−1 = ρ // 11 | 1.14581 | 1.03202 | 1.06531 | 1.09943 | 4.2187% // 13 | 1.00893 | 1.00809 | 0.98822 | 0.99622 | 1.2765% // 17 | 0.95486 | 0.97464 | 0.96921 | 0.94463 | 1.0827% // 19 | 0.92613 | 0.95626 | 0.95534 | 0.91355 | 1.3767% // 23 | 0.89305 | 0.94350 | 0.94311 | 0.88982 | 0.3625% // 29 | 0.87515 | 0.93424 | 0.93420 | 0.87277 | 0.2732% // 31 | 0.86259 | 0.92786 | 0.92784 | 0.86090 | 0.1963% // HL prediction with the q > y tail restored, δ_HL = ∏_{p>x}(1−1/(p−1)²) − 1: // @11: -2.6063% @13: -1.7900% @17: -1.3879% @19: -1.0800% @23: -0.8749% @29: -0.7482% @31: -0.6378% // measured δ minus HL-predicted δ (this difference IS the HL relative error at W = x#): // @11: 4.6728% @13: 1.3654% @17: 1.0891% @19: 1.3659% @23: 0.3600% @29: 0.2712% @31: 0.1950% // // === PART E — Suen (Janson 1998) applied, complete graph, anchored measure === // x | mu | Δ (unord) | δ=max_i Σ_{j~i}p_j | Δ* | Thm2 exponent | Thm3 bound | Thm8 bracket | truth S0/N̄ // 11 | 0.78909 | 0.2742 | 0.74654 | 0.628 | +0.4 | 0.83848 | -1.78e-1 | 0.500000 // 13 | 1.14676 | 0.6263 | 1.13520 | 2.038 | +4.9 | 0.84505 | -1.46e+1 | 0.310101 // 17 | 1.49379 | 1.0899 | 1.49097 | 4.987 | +20.0 | 0.84621 | -7.29e+2 | 0.208687 // 19 | 1.78566 | 1.5737 | 1.78502 | 9.587 | +54.1 | 0.84643 | -1.40e+5 | 0.152030 // 23 | 2.05259 | 2.0897 | 2.05246 | 16.559 | +124.7 | 0.84647 | -2.57e+8 | 0.112700 // 29 | 2.30539 | 2.6429 | 2.30537 | 26.900 | +263.5 | 0.84648 | -1.29e+13 | 0.085985 // 31 | 2.52361 | 3.1719 | 2.52360 | 40.069 | +491.0 | 0.84648 | -1.01e+19 | 0.068284 // Thm 2 exponent is positive at every level, so its bound exceeds 1 and says nothing. // Thm 3 is an UPPER bound on the survival probability, the wrong direction for // Assumption A, and it is loose by the ratio printed next: // @11: 1.68× @13: 2.73× @17: 4.05× @19: 5.57× @23: 7.51× @29: 9.84× @31: 12.40× // Thm 8 is the lower bound the programme needs. Its bracket 1 − Δ₀*e^{Δ*} is // positive only at @11 and negative from @13 on, by the magnitudes printed next // (a negative bracket means the theorem asserts P(S=0) >= a negative number): // @11: 10^-0.8 @13: 10^1.2 @17: 10^2.9 @19: 10^5.1 @23: 10^8.4 @29: 10^13.1 @31: 10^19.0 // // === PART D — the wall in local-lemma coordinates === // (i) pairwise horizon: graph empty iff q·q' ≤ H for all q,q' ≤ x, i.e. H ≥ x². θ = 2. // (ii) complete graph (set-wise hypothesis unverifiable): need max_c c·∏_{5≤p≤x}(1−c/p) ≥ 2. // (iii) admissible repair: |S| ≤ m with x^m ≤ H, i.e. Bonferroni depth m, i.e. Brun pure sieve. // x | Σ_{5≤p≤x}2/p | best c | max c∏(1−c/p) | (ii) feasible | m₀ (depth) | θ_pure = lnH/lnx // 5 | 0.4000 | 5.00 | 5.00000 | yes | 0 | 1.000 // 7 | 0.6857 | 2.50 | 1.25000 | NO | 2 | 2.176 // 11 | 0.8675 | 1.92 | 0.85832 | NO | 2 | 2.213 // 13 | 1.0214 | 1.68 | 0.71831 | NO | 2 | 2.418 // 17 | 1.1390 | 1.51 | 0.63031 | NO | 2 | 2.488 // 19 | 1.2443 | 1.41 | 0.57626 | NO | 2 | 2.750 // 23 | 1.3312 | 1.32 | 0.53495 | NO | 2 | 3.444 // 29 | 1.4002 | 1.26 | 0.50497 | NO | 4 | 3.322 // 101 | 1.9588 | 0.94 | 0.36180 | NO | 4 | 3.864 // 199 | 2.2314 | 0.84 | 0.31917 | NO | 6 | 4.672 // 227 | 2.2587 | 0.83 | 0.31541 | NO | 6 | 4.642 // 229 | 2.2674 | 0.82 | 0.31426 | NO | 6 | 4.661 // 439 | 2.4841 | 0.76 | 0.28795 | NO | 6 | 4.836 // 1009 | 2.7315 | 0.70 | 0.26292 | NO | 8 | 5.915 // 1000003 | 4.1080 | 0.48 | 0.17713 | NO | 12 | 9.442 // θ targets for comparison: pairwise horizon 2 (= the p² rule); best published // dimension-2 sifting limit β₂ = 4.26645028414864191641 (research/SEARCH-CONVENTIONS.md §4). // ============================================================================ // READINGS // // 1. THE INSTRUMENT AGREES WITH EVERYTHING IT SHOULD, AND EXTENDS IT ONE LEVEL. // PART A reproduces paper/anchored-note.md §3's S(19) = 38,380, // S(23) = 597,475, S(29) = 12,307,838 and S(31) = 283,449,187 to the unit, // and its L0 at @23 is // 1,784,710, the exact figure verify-cofactor-convolution.md §2 reports for // the repaired i = 0 cell. PART C's β, π_L and π_R reproduce that report's // §7 table to all five printed digits at all five of its levels. Nothing here // imports a number from natal-cap-35-x-multiplicity.js or from any // natal-cap-* file; the sieve is written from anchored-note.md §1's // definitions alone. // // 2. JANSON'S Δ IS ZERO, EXACTLY, AT EVERY LEVEL, AND THAT IS THE WHOLE STORY // ON THE PRODUCT MEASURE. The dependency graph of the 2K strike events is a // perfect matching, one edge per scour prime, and every edge is a mutual // exclusion because q | r and q | r+2 would force q | 2. So Δ = 0 and every // bound in the family collapses to the exact product ∏(1−2/q). PART B's // Δ_match column is that, level by level. There is no slack for an // inequality to occupy. // // 3. THE PRE-REGISTERED ANSWER ON Δ VERSUS THE FORCED SCALE IS CONFIRMED, AND // THE MISMATCH GROWS. Forcing one event per prime instead — the only reading // under which Δ is nonzero — gives Δ_complete = 0.5483 to 6.3438 across // @11..@31 against a forced scale F of 2.1086% down to 0.6419%, a ratio // running 26.0×, 73.3×, 157.3×, 289.3×, 473.9×, 701.3×, 988.3×. The // pre-registration said two to three orders of magnitude; at @11 it is 1.4 // orders, so PR2 is optimistic at the shallow end. Janson's e^{−Δ/2} is // a constant-factor instrument aimed at a sub-percent target, and the aim // gets worse with x, not better. // // 4. THE FORCED SCALE IS THE HARDY-LITTLEWOOD PREDICTION FOR δ, NOT A BOUND ON // IT. The identity (1+δ)(1+F) = 1+ρ is exact by construction, and the // δ_HL = −F/(1+F) column is what δ must become if ρ → 0. Measured δ minus // HL-predicted δ falls 4.6728%, 1.3654%, 1.0891%, 1.3659%, 0.3600%, 0.2712%, // 0.1950% across the seven levels. So δ is the forced scale with a MINUS sign plus a // residual that is decaying, and the report's reading of |δ| ≲ F as evidence // of near-independence has the sign of the main term backwards. // // 5. δ IS NOT BOUNDED BY THE FORCED SCALE. ρ/F = 2.001, 0.747, 0.781, 1.266, // 0.411, 0.362, 0.306. At @11 and @19 the residual exceeds the forced scale, and // since δ = ρ − F to second order that is exactly where δ comes out positive: // +2.0665% at @11 and +0.2859% at @19 against negatives elsewhere. // // 6. THE @19 SIGN FLIP NEEDS NO MECHANISM. δ is a difference of two positive // quantities of the same order, one of them (F) smooth and monotone // decreasing, the other (ρ) an error term measured at 4.2187%, 1.2765%, // 1.0827%, 1.3767%, 0.3625%, 0.2732%, 0.1963%. ρ rises from @17 to @19 and // crosses F; // that is the flip, and it is the second crossing, the first being between // @11 and @13. A sign-definite δ would have been the anomaly. // // 7. SUEN'S INEQUALITY, ACTUALLY APPLIED, IS VACUOUS FROM @13 ONWARD, AND THE // NUMBER IS 10^18.7 AT @31. PART E runs Janson (1998) Theorems 2, 3 and 8 on // the anchored measure, where Remark 3 forces the complete dependency graph. // Theorem 2's exponent is +0.4, +4.9, +20.0, +54.1, +124.7, +263.5, +491.0 — // positive at every level, so the bound exceeds 1. Theorem 3 gives 0.83848 to // 0.84648, an upper bound on survival that is loose by 1.68× to 12.40× and // points the wrong way for Assumption A anyway. Theorem 8, the lower bound the // programme actually needs, has bracket 1 − Δ₀*e^{Δ*} positive only at @11, // at 1.31e-1, and negative from @13 on at -1.03e+1, -4.88e+2, -8.68e+4, // -1.46e+8, -6.50e+12, -4.56e+18. The tool the brief named // first has exactly one level of content and it is the smallest one. // // 8. THE LOCAL LEMMA'S PAIRWISE HORIZON IS EXACTLY H = x², THE TARGET. Join // q ∼ q' iff a window of length H fails to equidistribute mod q·q'. At // H ≥ x² no pair of primes ≤ x is joined, the graph is empty, and the local // lemma would certify a survivor at H = x², which is the p² rule and the twin // prime conjecture. The step that fails is the quantifier: Erdős-Spencer's // lopsidependency asks for P(A_i | ∧_{j∈S} Ā_j) ≤ x_i ∏(1−x_j) for ALL // S ⊆ [n]∖(N(i)∪{i}), and verifying that needs equidistribution modulo // ∏_{q∈S} q, which at |S| = π(x) is x# and not x². // // 9. THE SET-WISE VERSION DIES AT x = 7, EARLIER THAN THE UNION BOUND. PART D's // (ii) column: max_c c·∏_{p≠q}(1−c/p) reads 5.00000, 1.25000, 0.85832, // 0.71831, 0.63031, 0.57626, 0.53495, 0.50497 at x = 5..29 against a required // 2. The Mertens sum Σ_{5≤p≤x}2/p crosses 1 between x = 11 and x = 13, which // is where sift-limit-attack.md §7's covering economy dies; the local lemma's // condition is strictly stronger and dies two primes earlier. This is the // same wall, reached from a new direction, and nothing in REFUTED.md's // covering-economy row is re-opened by it. // // 10. THE ADMISSIBLE REPAIR IS BRUN'S PURE SIEVE AND IT LANDS AT Θ(lnln x). // Truncating the hypothesis to sets with ∏_{q∈S}q ≤ H is Bonferroni at // depth m = ln H/ln x. The exact depths are 2, 2, 2, 2, 4, 4, 6, 6, 8, 12 at // x = 11..10⁶ and the exponents 2.213, 2.418, 2.488, 2.750, 3.444, 3.322, // 3.864, 4.672, 4.836, 5.915, 9.442. The crossing of the published // dimension-2 sifting limit happens between // x = 101 and x = 199 and never reverses. These are the same object // sift-limit-attack.md §7 tables as β_pure, computed from a different // starting point with a cruder error accounting, agreeing to about 10% and // diverging identically. // // 11. WHAT BLOCKS THE WALL IS NEITHER DEGREE NOR EXPECTED KILLS. Degree is 0 at // the pairwise horizon. The expected kill count μ = Σ2/q is 2.05259 at @23 // and 2.52361 at @31, comfortable for any of these theorems. The binding // quantity is the admissible conditioning-set modulus ∏_{q∈S} q ≤ H, which // caps |S| at ln H/ln x. Every exponent in reading 10 is that cap priced. // // 12. REACH. Seven levels, @11..@31, in 576.1 s, of which @31 (W = 2.0e11) is // nearly all. @37 is 37× @31 and was not attempted. PART D's LLL sweep runs to // x = 10⁶ and is arithmetic, not enumeration, so it carries no level ceiling // of its own. Every Suen quantity in PART E uses ν-values p_ip_j for the // pair expectations; the true anchored pair correlations differ by percents // and the vacuity margins are up to 10^18.7, so nothing turns on it. // --------------------------------------------------------------------------- // FIGURE PROVENANCE (added 2026-08-20, readings-traceability pass). One figure // here is not in the OUTPUT block. No number above was changed. // // ROUNDING of a value this file's own OUTPUT prints: W at @31 is printed in // full as 200560490130 and quoted as 2.0e11. // ---------------------------------------------------------------------------