# Red team: the Z2 foundation — natal-onset-01, zonegap-01, zonegap-prior-art *(2026-08-21. Adversarial pass over the three HELD artifacts the Z2 pivot builds on. Default refuted-until-rederived; independent code in the session scratchpad; page images for the prior-art spot checks. Verdicts CONFIRMED / WEAKENED / REFUTED per claim. Producer scripts were NOT edited; only this file is written.)* ## Verdict line, corrections first The foundation HOLDS. Two prose defects found, both WEAKENED-not-refuted; every load-bearing number and identity reached was rederived exactly by independent code (scratchpad: `rt-natal-verify.js`, `rt-natal-lens.js`, `rt-zonegap-sweep.js`, `rt-custody.js`; different sieve, different max-gap algorithm — block-max over a stored opener array, not the producer's deque). 1. **WEAKENED (zonegap-01.md §1): "3,909 rows parsed" is wrong — the TOS table parses to 4,051 rows under the producer's own regex** (rerun verbatim on `research/tos-twin-gaps-1e16.txt`, sha256 verified 78767cad...). The figure appears only in the staging prose, not in the embedded OUTPUT (whose in-code guard is `>= 3000` plus the 75-starred check); every custody consequence (75 starred rows, 1,116 shared sizes, 0 mismatches) is unaffected and was independently reproduced. Corrected sentence: "4,051 rows parsed". 2. **WEAKENED (natal-onset-01.md §4 wording): "Every level's record sits a factor 10^6-10^11 beyond its own x^2"** — true for the five comparison levels {29..43} (1.43e6..4.49e11), but the table's x = 23 row sits at 1.44e5. Corrected sentence: "each comparison level's record sits a factor 10^6-10^11 beyond its own x^2 (23's sits 1.4e5 beyond)". The clean-negative verdict is unaffected. ## 1. TARGET 1 — natal-onset-01: ALL FOUR CLAIMS CONFIRMED - **(a) q^2 ≡ {1,19} mod 30 mini-theorem: CONFIRMED, rederived over ALL residues.** The 8 residues coprime to 30 are {1,7,11,13,17,19,23,29}; their squares mod 30 are {1,19,1,19,19,1,19,1} — set {1,19}, both in the closer classes {13,19,1}; the openers q^2-2 fall in {29,17}, both in {11,17,29}. So the first fresh kill q^2 always lands B-side on the channel pair (q^2-2, q^2), and onset2 = onset1. Onset3 = q^2 iff q^2-2 prime: verified for all 22 scour primes (10 yes-rows; lags 88..1162 match the bound table row by row). - **(b) live A/B identity: CONFIRMED — it is a theorem of the model, with the suspected edge case worked and closed.** In the timeline model a member dies exactly once, at its own coordinate, killed by its lpf. An A-side kill strikes the opener a at time a; the partner's only possible death time is a+2 > a, so no smaller prime can have killed the B side FIRST — a strike at position a+2 happens at time a+2 regardless of which prime owns it. Hence every A-side channel kill destroys a live pair, and a B-side kill (time a+2) finds the pair dead iff the opener is composite (died at a < a+2), i.e. destroys a live pair iff the opener is prime. Checked mechanically by direct death-time simulation over all 796 destroyed pairs: 0 violations either side. The identity is exact, not approximate, and the "smaller prime" worry dissolves because kill TIME is the position, not the killer's size. - **(c) 1..500 timeline: CONFIRMED — independent sieve (trial division, no shared code) reproduces the bound OUTPUT exactly**: all 28 destruction rows (time, destroyer, pair, side, factorization), 22 live pairs at 500 (list identical), window totals 999/796/203, the classical cross-check 205 total twin pairs below 10^4, the A/B table per-q (kills, dA, dB, f_meas), aggregate 539:543, live A:B 2.10, B-pred 246.0, derived null 2.19, and the mirror conjugation mu(n) = (-2-n) mod 30q EXACT for all 22 primes both directions. Deepest-tread staircase row re-verified (58 pairs, 10 survive, 17.2%; prod(1-2/p) over 7..97 = 0.1915). - **(d) records-vs-x^2: CONFIRMED.** The six LADDER positions match `research/exact-g2-ladder.js` verbatim (incl. nmax = 2 at 37). Recomputed independently: pos/x#, ln pos/ln x#, pi(sqrt(pos)), pos/x^2, and the inter-prime-square fraction agree in every cell (37: q1 = 738173, pos - q1^2 = 107482, frac 0.0030), and all four ranks reproduce (4/5, 3/5, 3/5, 1/5). The x = 37 clean negative stands as stated, modulo the wording defect in the verdict line above. ## 2. TARGET 2 — zonegap-01: CONFIRMED at every reachable point - **(a) independent re-sieve, sample far beyond the brief: CONFIRMED.** Own sweep to X = 1e10 (`rt-zonegap-sweep.js`, 25 s): stores every twin opener and answers each zone by binary search + sqrt-decomposition block-max — a different algorithm from the producer's monotonic deque. Results: 27,412,679 pairs (the prereg's stage-1 figure exactly) over 9,591 zones (confirming the disclosed 9,592-in-prose miscount). Band rows 10^0..10^3 match the embedded 1e11 OUTPUT cell for cell (zones 143/1,061; c3 3.426±0.497 / 3.681±0.362; maxLoad 0.7504/0.7504; worst margin 1/5.00e1 / 1/7.06e2; head 0.942/0.684; meanU 0.623/0.672; frac u>0.8 0.315/0.308), and the 10^4 band matches the prereg's sha-bound stage-1 anchors (8,362 zones, 4.022±0.243, u 0.668, frac 0.274, head 0.720, load 0.7432, margin 1/3.49e4). All 34 envelope steps to 1e10 identical (positions AND u values); record ladder == A113274 1..34; worst head 0.7218 at p = 659 (head 150); stage-1 fit e = 3.411 and E-form 3.281 reproduce, and my OWN-RNG matched-noise controls read truth ln^3 as 3.000±0.008 and ln^2 as 2.000±0.007 — the calibration story is real, the estimator is unbiased on truth, so the drifting-exponent verdict (no single exponent quotable) is the right reading. Custody from primary sources (`rt-custody.js`): the inlined REC arrays == OEIS b113274/b113275 rows 1..41 EXACT (fetched fresh; record 42 starts 1.34e11 > X, so "records 1..41" is the complete set); the TOS file's 75 starred rows == the OEIS b-files EXACT; first-occurrence diff vs TOS to 1e9 both directions: 555 sizes, 0 mismatches; the envelope identity env(p) == largest record wholly below p'^2 re-verified EXACT at my own 3,400 zones to 1e9. maxgap-law §8: 7 of 7 reproduce (498..2868). - **(b) boundary conventions: CONFIRMED, no headline is convention-sensitive.** Gap is opener-to-opener — required, since that convention is what makes the ladder equal A113274 (verified: my ladder reproduces it); closer-to-opener would shift every gap by -2, no headline moves. Head from p, strict p < a: the only real choice is excluding a = p when p itself opens a twin pair; it changes head by O(1) at small p and touches Z2 never in the shared-level table. a+2 < p'^2 strict vs non-strict is vacuous (a+2 is prime, p'^2 is not). u from gapStart vs gapEnd tested side by side in-band: meanU 0.623 vs 0.629 (10^2), 0.668 vs 0.668 (10^4); frac u>0.8 identical to 3 decimals — the deep-end-loading claims survive both conventions. - **(c) margin ladder + G2 relation: CONFIRMED.** Worst Z2/width per band rederived: 1/3.50 (degenerate p = 2), 1/50.0, 1/706, 1/34,900 — the ladder's first four rungs exact; the top-band 1/1.66e6 was not re-swept (below) but is consistent with record 34 (6,282) against a ~1e10 width at the band's bottom. The shared-level table reproduces in every cell (widths, Z2, G2/Z2 = 1.00 at p = 2..7, then 1.40..4.12 at 43), and the one-line theorem is sound: in the zone, tile twin slots and twin-prime pairs coincide (certification both ways), so in-zone gaps are a subset of the tile's slot gaps and Z2(p) <= G2(p#). - **(d) every zone >= 2 pairs: CONFIRMED at the sample** — minimum count 2 (at p = 2), zero single-pair zones over all 9,591 zones to 1e10; the 1e11 statement rests on the embedded OUTPUT, whose every reachable number checked out. - **(e) prereg custody: CONFIRMED with one stated limit.** cdd753c (2026-08-21 13:49:01) adds exactly one file, the prereg, 63 lines, nothing else; the file is byte-identical at HEAD; the sweep's results commit is eb0f9c8 at 14:07:11 — 18 min later, consistent with the 315.7 s run happening after the seal. Git cannot PROVE the run did not happen pre-seal; corroborating the honest ordering: 3 of 5 predictions MISSED, and the misses share the record-correlation mechanism the doc names. Scoring re-checked against the sealed bands: P1 3.930 not in [4.05,4.35] MISS; P2 0.728 not in [0.63,0.71] and 0.462 not in [0.21,0.33] MISS; P3 0.725 in [0.64,0.80] PASS; P4 0.7272 in [0.66,0.80] PASS; P5 3.192 not in [3.38,3.48] and below the E-form 3.261 (sign flip) MISS — 2/5 exactly as reported. ## 3. TARGET 3 — zonegap-prior-art spot checks: ALL CONFIRMED at page image Fresh fetches this pass (scratchpad `rt-kw2019.pdf`, `rt-oehl-v1.pdf`, `rt-A192870.txt`); both PDF sha256s match the survey's §5 table exactly. - **(a) KW 2019: CONFIRMED.** Capital-T "Theorem" occurs 0 times in the pdftotext extraction; the 4 case-insensitive hits are "prime number theorem" (x3) and "Dirichlet's theorem". §2 heading is "Heuristics and Conjectures". The Generalized Cramér conjecture reads verbatim "Almost all maximal gaps Gc(p) satisfy Gc(p) < C_{k,H}^{-1} phi_{k,H}(q) log^{k+1} p" (their (20)); Generalized Shanks is the ~ form (21), also "almost all"; Gumbel and the O_k(log x) maximal-gap count are in the abstract; C_{2,H} ≈ 1.32032363169... sits in their appendix line (A005597, A114907) = 2C2 as claimed. - **(b) arXiv:1411.6582: CONFIRMED.** Abs page: v1 Thu 20 Nov 2014, v2 Sun 9 Aug 2015, "withdrawn by the author due to a crucial error in Lemma 8.4". v1 Conjecture 4.1 reads verbatim: "(Twin recursion conjecture). There exists at least one pair of twin primes {p̂, p̂ + 2} between the squares p2 and p'2 of two successive primes p and p'" (continuing "with distance Delta = p' - p such that p^2 < {p̂, p̂+2} < p'^2" — the survey's quote is a fair verbatim prefix). - **(c) A192870: CONFIRMED.** The TPC comment is verbatim ("there is no proof that the greatest integer M exists - not even for a subset of values of n. If one could find a constructive existence proof, then Twin Prime Conjecture as well as Legendre's Conjecture would require just a trivial additional step." — Pfoertner edit Sep 15 2021), the HL "indication (but not a proof)" comment is present, a(2) = 122, and the Raab a(6) > 3005845357 bound with the a(7)..a(10) estimates matches the survey's §5 note. One nit, sub-defect: the survey's %N paraphrase drops "of any possible pattern" and the "-1 if no such maximum" clause — content unchanged. ## NOT REACHED - **The 1e11 top band was not independently re-swept** (compute budget; producer run is 315 s and siblings are building). Everything specific to p in [1e5, 316228) — c3 3.930, meanU 0.728, frac 0.462, load 0.7272, margin 1/1.66e6, whole-sweep e = 3.192, head 0.7229, the 1,116-size first-occurrence count, 27,292-zone postulate minimum — rests on the embedded OUTPUT, against which every value reachable at X <= 1e10 (bands, envelope, anchors, estimator, controls) reproduced exactly. - First-occurrence diff vs TOS run to 1e9 (555 sizes both directions), not the full 1e11 range; TOS beyond its A113274-shared prefix remains single-witness, as the producer itself states. - Prereg RUN-order: git proves seal-alone-before-results-commit; that the 1e11 sweep did not execute pre-seal is corroborated (3/5 misses), not provable from the repository. - natal-onset: tread table verified at its deepest row + via the identical 1..500 timeline, not recomputed row by row; the lag distribution and A/B-null items the producer itself lists as NOT REACHED stay open. - Kourbatov 2013 not re-fetched (standing verbatim record `lit-pdf-kourbatov-grob.md` cited per the compute rule); Guy, Ribenboim, Korevaar, Wolf-1990s stand exactly as the survey's own NOT-REACHED says. - `node research/qc.js` after writing this file: TOTAL 3, all three in sibling files mid-flight (`zonegap-02-reduction.js`, `zonegap-03-model.js`: 2 uncited-script + 1 binding-lost) — not this pass's findings, left for their authors per the never-touch rule. --- *History layer: process record, staging. Independent code in the session scratchpad (`rt-natal-verify.js`, `rt-natal-lens.js`, `rt-zonegap-sweep.js` at 1e8/1e10, `rt-custody.js`); page images `rt-kw2019.pdf` sha256 7d69b356..., `rt-oehl-v1.pdf` sha256 f2e29048... See `research/history/CHANGELOG.md` for the corpus rule.*