# K* extended six steps: the drift is structural, the certificate reaches 18 of 19.2455 at the chain step itself, and the HL first moment becomes a period-free certificate engine *2026-08-21. Producer: `research/attack-kstar-01.js` (345 s, formal embed with input hashing; re-verify with `node research/qc/embed.js --check --timeout 900 research/attack-kstar-01.js` — the run walks the full 29# and 31# periods, so the default 120 s check timeout is not enough). Pre-registration: `attack-kstar-01-prereg.md`, committed ALONE at `552143c` before any of the three period walks ran — the producer's tile-only prediction sections existed and had run at commit time; the measurement sections did not exist yet. Answers `attack-doubling-01.md` §4's alarm and §6's "K* beyond 23# is unenumerated". HELD, not integrated — awaiting the standing one-pass adversarial review. Legend as in `research/sift-limit-attack.md`: **[PROVEN]** proof given here or by exact finite computation with validated engines; **[VERIFIED]** checked computationally here; **[MEASURED]** empirical, finite range; **[INFERRED]** deduction from sourced facts.* --- ## 0. The verdict, up front > **The alarm is structural, and it lands early. K* is now enumerated at 17 > doubling steps (was 11): three new full period walks (13#→29#, 13#→31#, > 17#→31#) and three period-free certificates past every scannable level > (19#→37#, 19#→41#, 19#→43#). The raw drift slope steepens from 0.6881 ± > 0.1328 to 0.8184 ± 0.0908, and the certificate touches K*+1 = 18 — 93.5% of > 2^β₂ = 19.2455 — at the base-2 chain step s = 16 itself, the sup row. > The instrument that did it: the kill-run count N_k is EXACTLY an > alternating sum of Hardy–Littlewood local products (W1's ν-case table one > level up), computable on the tile alone at cost D·2^k independent of the > period — so "expected count < 1" is not a heuristic here but an integer > reading N_k = 0, a proven finite-level certificate. The pre-registered > predictions hit at all three scanned steps, cell for cell (43 census > cells), and the HL growth verdict is NEAR-LINEAR: the bridging-certificate > route to an all-s C₂ < 19.2455 is dead — under the model it fails at the > very next chain step 31#→61#, and no fit is needed to see cert = 18 at > P(2s) = 31. C₂ itself stays flat (3.22–5.27 across the six new steps); > what dies is the certificate, not the inequality.** ## 1. (a) The extended ladder Three full cyclic walks (periods 6.47·10⁹ and 2.006·10¹¹, twice), each re-deriving the exact ladder from scratch as its custody gate — **G2(29#) = 258 @ 1205437109 (×2)** and **G2(31#) = 348 @ 8813641451 (×4)**, digit-for-digit including least argmax and multiplicity, alive counts = Π(p−2) — plus three certificate-only steps (§3). The fold-recursion path was also pushed from 13# to 23# (D₂₃ = 7,952,175), so ladder rows 17, 19, 23 are now re-derived by BOTH paths. **[VERIFIED]** | step | s | K* | cert = K*+1 | C₂ exact | cert/C₂ | route | |---|---|---|---|---|---|---| | 13#→29# | 15 | 10 | 11 | 43/11 = 3.9091 | 2.81 | scan | | 13#→31# | 16 (chain) | **17** | **18** | 58/11 = 5.2727 | 3.41 | scan | | 17#→31# | 17, 18 | 13 | 14 | 29/9 = 3.2222 | 4.34 | scan | | 19#→37# | 19, 20 | 13 | 14 | 528/150 = 3.5200 | 3.98 | I–E, period-free | | 19#→41# | 21 | 16 | 17 | 546/150 = 3.6400 | 4.67 | I–E, period-free | | 19#→43# | 22, 23 | 20 | 21 | 618/150 = 4.1200 | 5.10 | pruned covering DP (kstar_dfs.py, #1144), period-free | Full ladder (K*+1, 17 steps): 3 2 5 3 4 4 6 9 7 11 9 11 18 14 14 17 21. All five of this pass's new C₂ comparisons are custody-exact (both G₂ terms among the 14 exact ladder terms); the sixth, 19#→43#'s, is exact on the same footing by its own report (#1144). The certificate holds at every step and its slack cert/C₂ **widens monotonically in scale**: 1.00 at the first three steps (attack-doubling-01) to 5.10 at 19#→43#. **[VERIFIED]** ## 2. (b) The HL route, made exact — and the growth verdict **The identity [PROVEN, elementary].** A slot x survives the entering primes Q iff t + x avoids {0, −2} mod every q ∈ Q; for a window of k consecutive level-s slots x₁..x_k, inclusion–exclusion over the surviving subset J gives, per period of P(2s)#, > N_k = Σ_shapes Σ_{J⊆[k]} (−1)^{|J|} Π_{q∈Q} (q − ν_q(J)), > ν_q(J) = #distinct residues of {x_j, x_j+2 : j ∈ J} mod q — an INTEGER identity, no probability anywhere. Each ν_q(J) is the Hardy–Littlewood local count of the 2|J|-tuple {x_j, x_j+2}: on pairs it is exactly W1's case table (ν = 4 − 2[q|d] − [q|d−2] − [q|d+2], verified on 295 cells), so a kill-run IS a k-fold twin-tuple correlation event and N_k is the HL k-tuple first moment with nothing truncated — the `w1-singular-series.md` identification one level up. K* = max{k : N_k ≥ 1}, and N_k = 0 is a proven certificate. Cost: D_s·2^k·|Q| with monotone shape-pruning (an all-killed (k+1)-window needs both its k-sub-windows all-killed), INDEPENDENT of the period. **[PROVEN + VERIFIED: the I–E engine reproduces the scan's full run census at every k at all 14 scanned steps — 63 (step, k) cells at the eleven known steps, 43 more at the three new ones, and K* equal at all 14.]** **The model scorecard [MEASURED].** Two smooth rungs against the exact count, eleven known steps: M0 (iid: N_k = D·M·p₁^k, p₁ = 1 − Π(1−2/q)) overshoots K* by **+4.02 mean**; M1 (pairwise singular-series/Kirkwood: per-shape Π_{j