// ============================================================================ // FOLD PROFILE 13 — THE SLIDING SWEEP: closing the coverage gaps of 12 // ============================================================================ // Chris, 2026-08-17: did we check ranges around the landmarks, or only the exact // sets? Answer: ranges, at several scales, plus a 44,607-window dyadic sweep. // But that sweep had three real gaps, and this file closes all three. // // GAP 1 the dyadic sweep used ALIGNED windows (offsets 0, w, 2w, ...), so a // hotspot straddling a boundary would be diluted. Fixed here with a // SLIDING sweep, step = one bin, at four widths. // GAP 2 only T23 was tested. Fixed here with T17, T19 and T23. // GAP 3 the seams had a single control offset each. Fixed here with 400 // random controls per width, giving a null distribution and a z-score. // // AND THE STATISTICAL POINT 12 DID NOT MAKE. With tens of thousands of windows, // some will beat the mean by chance. The honest test is whether the extremes // exceed what that many draws should produce. So every window is scored against // a LOCAL baseline (the survival of a 10x wider window on the same centre), which // removes the ln^2 y trend, and the extreme z is compared with sqrt(2 ln N), the // expected maximum of N standard normals. // // Method: bin each tile into ~200k bins, count slots and twin primes per bin, // prefix-sum, then every window is O(1). One primality sieve serves all tiles. // // Run: node --max-old-space-size=7000 fold-profile-13-hotspot-sweep.js // ============================================================================ 'use strict'; const t0 = Date.now(); const el = () => ((Date.now() - t0) / 1000).toFixed(1) + 's'; const log = (s) => process.stderr.write(' [' + el() + '] ' + s + '\n'); const TILES = [ { x: 17, primes: [2, 3, 5, 7, 11, 13, 17] }, { x: 19, primes: [2, 3, 5, 7, 11, 13, 17, 19] }, { x: 23, primes: [2, 3, 5, 7, 11, 13, 17, 19, 23] }, ]; for (const T of TILES) T.W = T.primes.reduce((a, b) => a * b, 1); const MAXW = Math.max(...TILES.map((T) => T.W)); log(`sieving primes to ${(MAXW + 2).toExponential(2)} ...`); const comp = new Uint8Array(MAXW + 3); comp[0] = comp[1] = 1; for (let i = 2; i * i <= MAXW + 2; i++) if (!comp[i]) for (let j = i * i; j <= MAXW + 2; j += i) comp[j] = 1; console.log('='.repeat(100)); console.log('FOLD PROFILE 13 — sliding sweep for a local twin-survival hotspot, three tiles'); console.log('='.repeat(100)); function buildTile(T) { const { W, primes } = T; const rough = new Uint8Array(W + 3).fill(1); rough[0] = 0; for (const q of primes) for (let m = q; m <= W + 2; m += q) rough[m] = 0; const NB = 200000; const B = Math.ceil(W / NB); const nb = Math.ceil(W / B); const bs = new Int32Array(nb + 1), bt = new Int32Array(nb + 1); for (let r = 0; r < W; r++) { const b = (r / B) | 0; if (rough[r] && rough[r + 2]) bs[b]++; if (!comp[r] && !comp[r + 2]) bt[b]++; } const PS = new Int32Array(nb + 1), PT = new Int32Array(nb + 1); for (let i = 0; i < nb; i++) { PS[i + 1] = PS[i] + bs[i]; PT[i + 1] = PT[i] + bt[i]; } return { W, B, nb, PS, PT, D: PS[nb], TT: PT[nb], rough }; } const results = []; // GROUPING IS PINNED to en-US (2026-09-24). A bare toLocaleString() follows the // HOST locale, so on a French machine (510,510) prints "510 510" and this file's // embedded OUTPUT block -- captured on a comma-grouping host -- cannot reproduce // there; `research/qc/embed.js --check` then reports "this run does not reproduce // the pasted block" on a file that has not changed. These four figures are the // only locale-dependent bytes this script prints. for (const T of TILES) { log(`building T${T.x} (W = ${T.W.toLocaleString('en-US')}) ...`); const G = buildTile(T); const MEAN = G.TT / G.D; console.log(''); console.log(`T${T.x}: W = ${G.W.toLocaleString('en-US')}, bin = ${G.B}, slots D = ${G.D.toLocaleString('en-US')}, ` + `twin primes = ${G.TT.toLocaleString('en-US')}, mean survival = ${MEAN.toFixed(6)}`); console.log('-'.repeat(100)); const S = (i, j) => G.PS[Math.min(j, G.nb)] - G.PS[Math.max(i, 0)]; const Tw = (i, j) => G.PT[Math.min(j, G.nb)] - G.PT[Math.max(i, 0)]; // Detrending: the baseline is the mean of the two FLANKING windows of the same // width, which removes a smooth gradient. The 10x-wider baseline used first // failed at the origin, where survival changes by a factor across the baseline // itself, and it manufactured every large z it reported. // The head is also excluded below, since its gradient is the known // crystallisation decline and is not what we are hunting. console.log(' width (bins) | windows | z sd | min z | max z | |z|>4 | threshold | top extremes (position of W : z)'); for (const wb of [8, 80, 800, 8000]) { if (wb * 6 > G.nb) continue; // Both flanks must FIT: a clipped flank destroys the symmetry that cancels the // trend, and that alone produced every large z in the first version of this sweep. const HEADCUT = Math.max(Math.ceil(G.nb / 1000), wb); let n = 0, mn = Infinity, mx = -Infinity, big = 0, sum = 0, sum2 = 0; const top = []; for (let i = HEADCUT; i + 2 * wb <= G.nb; i++) { const s = S(i, i + wb); if (s < 100) continue; const t = Tw(i, i + wb); const ls = S(i - wb, i), lt = Tw(i - wb, i); const rs = S(i + wb, i + 2 * wb), rt = Tw(i + wb, i + 2 * wb); const fs = ls + rs, ft = lt + rt; if (fs < 100) continue; const p = ft / fs; if (p <= 0 || p >= 1) continue; const mu = s * p, sd = Math.sqrt(s * p * (1 - p) * (1 + s / fs)); const z = (t - mu) / sd; n++; sum += z; sum2 += z * z; if (z < mn) mn = z; if (z > mx) mx = z; if (Math.abs(z) > 4) big++; top.push({ z: Math.abs(z), sz: z, pos: (i * G.B) / G.W }); } if (!n) continue; const zmean = sum / n, zsd = Math.sqrt(sum2 / n - zmean * zmean); top.sort((a, b) => b.z - a.z); const thr = zsd * Math.sqrt(2 * Math.log(n)); // scaled by the MEASURED dispersion const head = top.slice(0, 3).map((e) => e.pos.toFixed(3) + ':' + e.sz.toFixed(1)).join(' '); console.log(` ${String(wb).padStart(12)} | ${String(n).padStart(8)} | ${zsd.toFixed(2).padStart(5)} | ` + `${mn.toFixed(2).padStart(6)} | ${mx.toFixed(2).padStart(5)} | ${String(big).padStart(6)} | ` + `${thr.toFixed(2).padStart(10)} | ${head}`); results.push({ x: T.x, wb, n, mn, mx, big, expMax: thr, zsd }); } // ---- seams with a proper null ------------------------------------------- if (T.x >= 19) { const WPREV = G.W / T.x; console.log(''); console.log(` SEAMS at multiples of ${WPREV.toLocaleString('en-US')}, k = 1..${T.x - 1}, against 400 random controls each:`); console.log(' half-width | seam survival | control mean | control sd | z of the seams | verdict'); let seed = 12345; // Math.imul, because (seed * 1103515245) reaches 2.37e18 = 263 x 2^53 and // is inexact from the second call: the plain-multiply form collapses to a // period of 10,466 with a tail of ~5,900, against ~2^31 for the real thing, // and its low three bits sit in one bucket 99.6% of the time. The imul form // reproduces the exact LCG mod 2^31 step for step. (2026-08-20.) const rnd = () => { seed = (Math.imul(seed, 1103515245) + 12345) & 0x7fffffff; return seed / 0x7fffffff; }; for (const h of [3000, 30000, 300000]) { const hb = Math.max(1, Math.round(h / G.B)); let ss = 0, st = 0; for (let k = 1; k < T.x; k++) { const c = Math.round((k * WPREV) / G.B); ss += S(c - hb, c + hb); st += Tw(c - hb, c + hb); } const seamR = st / ss; const ctrl = []; for (let trial = 0; trial < 400; trial++) { let cs = 0, ct = 0; for (let k = 1; k < T.x; k++) { const c = Math.floor(rnd() * (G.nb - 2 * hb)) + hb; cs += S(c - hb, c + hb); ct += Tw(c - hb, c + hb); } ctrl.push(ct / cs); } const cm = ctrl.reduce((a, b) => a + b, 0) / ctrl.length; const csd = Math.sqrt(ctrl.reduce((a, b) => a + (b - cm) * (b - cm), 0) / ctrl.length); const z = (seamR - cm) / csd; console.log(` ${String(h).padStart(10)} | ${seamR.toFixed(5).padStart(14)} | ${cm.toFixed(5).padStart(13)} | ` + `${csd.toFixed(5).padStart(11)} | ${z.toFixed(2).padStart(15)} | ${Math.abs(z) > 3 ? '*** ANOMALY ***' : 'consistent with the null'}`); } } } console.log(''); console.log('='.repeat(100)); console.log('SUMMARY — the hotspot hunt is a ONE-SIDED question: is any window ABOVE its neighbours?'); console.log('='.repeat(100)); console.log(' tile | width | windows | dispersion | max z (the hunt) | threshold | verdict'); let anyPos = false; for (const r of results) { const flag = r.mx > r.expMax; if (flag) anyPos = true; console.log(` T${String(r.x).padEnd(3)} | ${String(r.wb).padStart(5)} | ${String(r.n).padStart(8)} | ` + `${r.zsd.toFixed(2).padStart(11)} | ${r.mx.toFixed(2).padStart(17)} | ${r.expMax.toFixed(2).padStart(9)} | ` + `${flag ? '*** ABOVE ***' : 'nothing'}`); } console.log(''); console.log(anyPos ? ' A positive excess survived. Investigate.' : ' NO POSITIVE EXCESS at any width in any tile. The largest upward deviation anywhere\n' + ' is 4.90 sigma against a 4.75 threshold at 199,690 windows, and its three top\n' + ' extremes sit at 0.130, 0.907 and 0.967 of W with no clustering. That is noise.'); console.log(''); console.log(' TWO THINGS THE SWEEP FOUND ON THE WAY, both worth keeping:'); console.log(''); console.log(' 1. Every downward exceedance sits at 0.004 or 0.040 of W, i.e. still inside the'); console.log(' head. Symmetric flanking windows cancel a linear trend but not curvature, and'); console.log(' the crystallisation decline is steepest there. It is the known gradient, not a'); console.log(' cold spot. Outside the first few percent of the tile nothing exceeds either way.'); console.log(''); console.log(' 2. The dispersion of z is 0.84 to 0.96 at widths 8, 80 and 800, i.e. SUB-BINOMIAL.'); console.log(' Twin primes in the tile fluctuate LESS than independent sampling would give.'); console.log(' (The 8000-bin rows sit outside that band, at 1.38 with 4,273 exceedances at'); console.log(' T23; those windows are a large fraction of the tile and the flank baseline'); console.log(' is not trustworthy there. Read the deficit off widths 8 to 800 only.)'); console.log(''); console.log(' *** CORRECTED 2026-08-17. This reading originally called the deficit "an'); console.log(' independent sighting of the anchored calm recorded in NATAL-CAP-CAMPAIGN'); console.log(' (sub-random house splits, sub-CRT overlaps, anchored variance at 0.34-0.62x'); console.log(' ensemble), now seen in a fourth place and by a fifth method". That'); console.log(' identification was never tested and is now superseded twice over.'); console.log(' fold-profile-14 established the deficit is real rather than an estimator'); console.log(' artifact (seeded synthetic control 1.009, 0.994, 0.960 against real 0.962,'); console.log(' 0.921, 0.838),'); console.log(' and fold-profile-16 established it is NOT a tile phenomenon: it survives'); console.log(' with the tile normalisation removed entirely, so it is the ordinary'); console.log(' short-interval variance of twin primes, which is classical and expected.'); console.log(' The anchored calm is a statement about slot dynamics and ensemble variance,'); console.log(' a different measurement. Nothing here supports joining them. ***'); console.log(''); console.log(' done in ' + el()); // ============================================================================ // OUTPUT — EMBEDDED, do not hand-edit. Regenerate: // node research/qc/embed.js --streams both research/fold-profile-13-hotspot-sweep.js // invocation: node research/fold-profile-13-hotspot-sweep.js // code-sha256: 992248affe288782c009070eb152171aa4ce71582bb7fd5a91c269b3a07b0b17 // out-sha256: 2f869b3a66f5b4b4727893d4197c3db3ab8f01bca37029a8fd431a33322708eb // body-lines: 86 // streams: stdout+stderr // node: v24.18.0 // embedded: 2026-09-24 // elapsed: 4.7 s // ============================================================================ // ==================================================================================================== // FOLD PROFILE 13 — sliding sweep for a local twin-survival hotspot, three tiles // ==================================================================================================== // // T17: W = 510,510, bin = 3, slots D = 22,275, twin primes = 4,636, mean survival = 0.208126 // ---------------------------------------------------------------------------------------------------- // width (bins) | windows | z sd | min z | max z | |z|>4 | threshold | top extremes (position of W : z) // 800 | 166103 | 0.92 | -3.02 | 3.90 | 0 | 4.51 | 0.907:3.9 0.907:3.9 0.907:3.8 // 8000 | 146171 | 0.88 | -2.92 | 3.20 | 0 | 4.30 | 0.604:3.2 0.604:3.2 0.604:3.2 // // T19: W = 9,699,690, bin = 49, slots D = 378,675, twin primes = 57,453, mean survival = 0.151721 // ---------------------------------------------------------------------------------------------------- // width (bins) | windows | z sd | min z | max z | |z|>4 | threshold | top extremes (position of W : z) // 80 | 197596 | 0.95 | -3.54 | 4.52 | 15 | 4.68 | 0.470:4.5 0.470:4.4 0.470:4.3 // 800 | 195554 | 0.90 | -3.96 | 3.22 | 0 | 4.43 | 0.004:-4.0 0.004:-3.9 0.004:-3.9 // 8000 | 173954 | 0.87 | -5.46 | 2.03 | 907 | 4.27 | 0.040:-5.5 0.040:-5.4 0.040:-5.4 // // SEAMS at multiples of 510,510, k = 1..18, against 400 random controls each: // half-width | seam survival | control mean | control sd | z of the seams | verdict // 3000 | 0.14067 | 0.15154 | 0.00766 | -1.42 | consistent with the null // 30000 | 0.15057 | 0.15097 | 0.00579 | -0.07 | consistent with the null // 300000 | 0.14904 | 0.14958 | 0.00422 | -0.13 | consistent with the null // // T23: W = 223,092,870, bin = 1116, slots D = 7,952,175, twin primes = 896,062, mean survival = 0.112681 // ---------------------------------------------------------------------------------------------------- // width (bins) | windows | z sd | min z | max z | |z|>4 | threshold | top extremes (position of W : z) // 8 | 199690 | 0.96 | -3.80 | 4.90 | 15 | 4.75 | 0.130:4.9 0.907:4.6 0.967:4.5 // 80 | 199546 | 0.92 | -3.86 | 3.94 | 0 | 4.55 | 0.490:3.9 0.547:3.9 0.490:-3.9 // 800 | 197506 | 0.84 | -4.85 | 2.85 | 9 | 4.14 | 0.004:-4.9 0.004:-4.6 0.004:-4.6 // 8000 | 175906 | 1.38 | -13.77 | 1.65 | 4273 | 6.77 | 0.040:-13.8 0.040:-13.7 0.040:-13.7 // // SEAMS at multiples of 9,699,690, k = 1..22, against 400 random controls each: // half-width | seam survival | control mean | control sd | z of the seams | verdict // 3000 | 0.10910 | 0.11298 | 0.00569 | -0.68 | consistent with the null // 30000 | 0.11345 | 0.11259 | 0.00330 | 0.26 | consistent with the null // 300000 | 0.11113 | 0.11278 | 0.00310 | -0.53 | consistent with the null // // ==================================================================================================== // SUMMARY — the hotspot hunt is a ONE-SIDED question: is any window ABOVE its neighbours? // ==================================================================================================== // tile | width | windows | dispersion | max z (the hunt) | threshold | verdict // T17 | 800 | 166103 | 0.92 | 3.90 | 4.51 | nothing // T17 | 8000 | 146171 | 0.88 | 3.20 | 4.30 | nothing // T19 | 80 | 197596 | 0.95 | 4.52 | 4.68 | nothing // T19 | 800 | 195554 | 0.90 | 3.22 | 4.43 | nothing // T19 | 8000 | 173954 | 0.87 | 2.03 | 4.27 | nothing // T23 | 8 | 199690 | 0.96 | 4.90 | 4.75 | *** ABOVE *** // T23 | 80 | 199546 | 0.92 | 3.94 | 4.55 | nothing // T23 | 800 | 197506 | 0.84 | 2.85 | 4.14 | nothing // T23 | 8000 | 175906 | 1.38 | 1.65 | 6.77 | nothing // // A positive excess survived. Investigate. // // TWO THINGS THE SWEEP FOUND ON THE WAY, both worth keeping: // // 1. Every downward exceedance sits at 0.004 or 0.040 of W, i.e. still inside the // head. Symmetric flanking windows cancel a linear trend but not curvature, and // the crystallisation decline is steepest there. It is the known gradient, not a // cold spot. Outside the first few percent of the tile nothing exceeds either way. // // 2. The dispersion of z is 0.84 to 0.96 at widths 8, 80 and 800, i.e. SUB-BINOMIAL. // Twin primes in the tile fluctuate LESS than independent sampling would give. // (The 8000-bin rows sit outside that band, at 1.38 with 4,273 exceedances at // T23; those windows are a large fraction of the tile and the flank baseline // is not trustworthy there. Read the deficit off widths 8 to 800 only.) // // *** CORRECTED 2026-08-17. This reading originally called the deficit "an // independent sighting of the anchored calm recorded in NATAL-CAP-CAMPAIGN // (sub-random house splits, sub-CRT overlaps, anchored variance at 0.34-0.62x // ensemble), now seen in a fourth place and by a fifth method". That // identification was never tested and is now superseded twice over. // fold-profile-14 established the deficit is real rather than an estimator // artifact (seeded synthetic control 1.009, 0.994, 0.960 against real 0.962, // 0.921, 0.838), // and fold-profile-16 established it is NOT a tile phenomenon: it survives // with the tile normalisation removed entirely, so it is the ordinary // short-interval variance of twin primes, which is classical and expected. // The anchored calm is a statement about slot dynamics and ensemble variance, // a different measurement. Nothing here supports joining them. *** // // done in 4.6s // ───── stderr ───── // [0.0s] sieving primes to 2.23e+8 ... // [2.1s] building T17 (W = 510,510) ... // [2.3s] building T19 (W = 9,699,690) ... // [2.8s] building T23 (W = 223,092,870) ... // ============================================================================ // READINGS (2026-08-18) — honestly calibrated // ============================================================================ // 1. THE "*** ABOVE ***" IS NOTHING, AND HERE IS THE TEST THAT SAYS SO. The // summary's one positive, T23 at width 8 with max z = 4.90 against a // threshold of 4.75, closes with "A positive excess survived. Investigate." // Nobody did (wave 5 logged that as its open loose end S2-22). Investigated // here by simulation. Null model: twin primes placed INDEPENDENTLY at rate // p = 0.112681 in each slot, windows of s = 318 slots (8 bins x 1,116 // positions at the measured slot density) with the file's own flank // estimator and its own sd formula, sampled exactly. Result over 60 // replicates, now produced and embedded by // `research/fold-profile-13-null-mc.js` (see the note at the foot of this // file; these figures are that script's block, not this one's): // effective windows 24,962 (n / width, the correlated-field count): // max z median 4.78, p10 4.44, p90 5.51; 42% of replicates >= 4.90 // n = 199,690 as the file counts them: // max z median 5.48, p10 5.05, p90 6.08; 90% of replicates >= 4.90 // So a maximum of 4.90 is the TYPICAL outcome of this sweep under pure // independence, not an excess. Consider the excess retired. // 2. THE THRESHOLD IS THE BUG, AND IT IS A SKEW PROBLEM, NOT A COUNTING ONE. // The code computes `thr = zsd * Math.sqrt(2 * Math.log(n))`, i.e. the // expected maximum of n standard NORMALS rescaled by the measured // dispersion. Measured on the null above, the z statistic has sd 1.007 — // the rescaling is fine — but skewness +0.312, because at width 8 a window // holds only about 36 twin primes (= 318 slots x 0.112681) and a binomial count that small has a // heavy right tail. sqrt(2 ln n) knows nothing about that. The threshold is // therefore too low at the smallest width, which is exactly the width where // the one positive appeared. At widths 80, 800 and 8000 the counts are 10x, // 100x and 1000x larger, the skew is negligible, and the threshold is sound. // 3. WHAT THE SWEEP GENUINELY ESTABLISHES IS THE NEGATIVE, AND IT IS CLEAN. // Nine tile-width combinations, 1.65 million windows in total // (= the summary table's windows column summed), sliding at // one-bin steps so GAP 1 is genuinely closed, and after reading 1 there is // not one upward excursion beyond expectation anywhere. Every downward // exceedance sits at 0.004 or 0.040 of W, inside the head, and is the // crystallisation gradient the file names. There is no twin-survival hotspot // in these tiles at any scale from 9e3 to 9e6 positions (widths 8 and 8000 bins x 1,116 at T23). // 4. GAP 3 IS CLOSED AND THE ANSWER IS FLAT. Seams against 400 random controls // give z = -1.42, -0.07, -0.13 at T19 and -0.68, +0.26, -0.53 at T23, all // "consistent with the null". This is the file that owns those z-scores, and // the quoted range fuses two tiles: -1.42 is T19 at half-width 3,000 and // +0.26 is T23 at half-width 30,000, so it is not a range within any one // measurement. The six digits this reading carried before 2026-08-20 and why // they moved, with the verdict unmoved, are in the FIGURE PROVENANCE note // below. Wave 5 (S2-7) found research/GLOSSARY.md and // paper/moire-primes.md crediting both the 400 controls and the z range to // fold-profile-12, which prints no z at all. Checked today: both documents // have since been repaired and now name this file and flag the two-tile // fusion. S2-7 is CLOSED. // 5. GAP 2 IS CLOSED UNEVENLY AND THE SUMMARY TABLE HIDES IT. T17 gets two // widths and no seam block, T19 three widths, T23 four. The missing rows are // not an oversight: the bin size scales with W (3, 49, 1,116 positions), so // a width of 8 bins is 24 positions at T17 and holds about one slot, and the // `s < 100` guard discards every such window. The consequence is that the // column headed "width" is in BINS, and the same number means 2,400 // positions at T17, 39,200 at T19 and 892,800 at T23. The summary table // lists all nine rows together as if the widths were comparable. They are // not, and no cross-tile statement should be read off it. // 6. THE SUB-BINOMIAL DISPERSION IS REAL AND THE FILE'S OWN CAVEAT IS THE RIGHT // ONE. z sd reads 0.92, 0.88 (T17), 0.95, 0.90, 0.87 (T19), 0.96, 0.92, 0.84 // (T23) at widths 8 to 800, against a null sd measured at 1.013 // (`fold-profile-13-null-mc.js`). The 8000-bin // rows sit outside at 1.38 with 4,273 exceedances, and the file says so and // says why (the flank baseline is untrustworthy when the window is a large // fraction of the tile). fold-profile-14 then confirmed the deficit is not an // estimator artifact and fold-profile-16 established it is the classical // short-interval variance of twin primes and NOT a tile phenomenon. The // 2026-08-17 correction block above carries all of that; it is the model of // the form and nothing in this run contradicts it. // 7. SCOPE. Three tiles, all fully sieved, so nothing is estimated. The // detrending baseline is the two flanking windows of the same width, chosen // after a 10x-wider baseline was found to manufacture large z at the origin — // that history is in the code comment at the top of the sweep loop and is // worth preserving. The head is cut at nb/1000, so nothing below 0.1% of W // is tested at all. Runtime 2.6 s, plain node; the header's // --max-old-space-size=7000 was not needed. // // --------------------------------------------------------------------------- // FIGURE PROVENANCE (added 2026-08-20, readings-traceability pass). What every // figure in the readings above that the OUTPUT block does not contain // verbatim actually is. No number above was changed. // // DERIVED IN THIS READING by arithmetic over printed values: // s = 318 slots (reading 1) is 8 bins x 1,116 positions carried through the // printed T23 slot density, 8,928 x 7,952,175/223,092,870 = 318.24. // effective windows 25,000 (reading 1) is the printed T23 width-8 window // count over the width, 199,690/8 = 24,961, rounded. // 2,400 / 39,200 / 892,800 positions (reading 5) are 800 bins times the // printed bin sizes 3, 49 and 1,116. // IN-CODE: the `s < 100` guard is line 98 (and its flank twin at 103); the // head cut nb/1000 is the HEADCUT at line 93; --max-old-space-size=7000 is // the Run: line of the header at line 25. // RESOLVED 2026-08-20 (mismatch adjudication #32), by writing the missing // producer. The null-model figures of readings 1, 2 and 6 came from a // "scratch Monte Carlo, not this script" that was never pasted anywhere in // the corpus, and they are the ENTIRE basis for retiring this sweep's one // positive. That simulation now exists as // `research/fold-profile-13-null-mc.js`, with its own embedded OUTPUT: it // builds the real T23 slot geometry (asserting its slot count against the // 7,952,175 this file prints, before doing anything else), sprinkles twins // independently at the measured p = 0.112681, and runs THIS file's scoring // code path verbatim -- HEADCUT, the s >= 100 and fs >= 100 filters, the // flanking-window baseline, the (1 + s/fs) inflation. Readings 1, 2 and 6 // now quote its printed block. The reproduction confirms the scratch run at // every figure, with the small shifts a different RNG gives at 60 // replicates: // old (scratch, unsourced) new (printed by the companion) // 25,000 effective windows 24,962 effective windows // median 4.70 / p10 4.28 / median 4.78 / p10 4.44 / // p90 5.30 / 33% >= 4.90 p90 5.51 / 42% >= 4.90 // n = 199,690 n = 199,690 // median 5.38 / p10 5.01 / median 5.48 / p10 5.05 / // p90 5.95 / 95% >= 4.90 p90 6.08 / 90% >= 4.90 // z sd 1.007, skew +0.312 z sd 1.013, skew +0.233 // The verdict is unchanged and now rests on a block: max z = 4.90 sits at // the median of the null, so it is not an excess. // CORRECTED in the same pass (mismatch adjudication #18): reading 2 said the // counts at widths 80, 800 and 8000 are "10x to 10,000x larger" and now says // "10x, 100x and 1000x". Against width 8 those three widths give exactly // 10x, 100x and 1000x; the top of the old range overshot by a factor of ten. // The direction of the argument, that the skew is negligible away from // width 8, is unaffected. // RETIRED 2026-09-24, AND NOT A DIFFERENT DRAW (reading 4's six seam z scores, // the figures the audit of return #1074 caught). Reading 4 printed // z = -1.41, -0.19, -0.09 at T19 and -0.65, +0.13, -0.50 at T23 until this // pass. They are the exact printed output of the seam loop's pre-2026-08-20 // plain-multiply LCG, `seed = (seed * 1103515245 + 12345) % 2**31` -- inexact // from the second call, because that product passes 2**53 -- and NOT a // different random draw: restoring that one line in the seam loop reproduces // all six digits exactly, at the two decimals it prints. Today's rnd // (`Math.imul`, 2026-08-20) is deterministic, and it prints the table's // -1.42, -0.07, -0.13 and -0.68, +0.26, -0.53, which is what reading 4 now // quotes. Nothing in the reading moves but the digits: both LCGs put every // seam inside 3 sd of the null, which is the whole claim, and the two-tile // fusion the reading flags -- 3,000 in one tile, 30,000 in the other -- holds // under either. // ---------------------------------------------------------------------------