============================================================================== PART 0 — CUSTODY: the ladder, the copy theorem, the 6p sum rule ============================================================================== ------------------------------------------------------------------------------ 0.1 D and G2 ladders, from a generator that never sieves ------------------------------------------------------------------------------ ok T_5: D = 3 G2 = 12 ok T_7: D = 15 G2 = 30 ok T_11: D = 135 G2 = 42 ok T_13: D = 1485 G2 = 66 ok T_17: D = 22275 G2 = 108 ok T_19: D = 378675 G2 = 150 ok T_23: D = 7952175 G2 = 204 ------------------------------------------------------------------------------ 0.2 copy theorem, spot-check by this file's own machinery ------------------------------------------------------------------------------ (the full 40/40 family table is CITED, not recomputed: research/attack-0c0e-01-deleted-family.js, embedded 2026-08-19, max_a Delta_m = maxsum_m(T_p) for all m <= 8 at folds 11..31.) ok max_a maxsum_m(T_11 \ {a,a-2} mod 13) = maxsum_m(T_13) m1:66/66 m2:96/96 m3:138/138 m4:156/156 ok max_a maxsum_m(T_17 \ {a,a-2} mod 19) = maxsum_m(T_19) m1:150/150 m2:186/186 m3:210/210 m4:228/228 ------------------------------------------------------------------------------ 0.3 the 6p sum rule (A5 Theorem A) and the Alternation Lemma ------------------------------------------------------------------------------ qualifying gaps at fold q: g = 0 or +-2 (mod q). Any two ADJACENT gaps of a kill run lie in classes that strictly alternate, so their sum is at least min(class +2) + min(class -2) = 6q exactly. x q L #qual gaps #adj pairs #run-legal min pair sum min LEGAL sum 6q attained? 7 11 1 0 0 0 - - 66 no legal pair 11 13 2 6 0 0 - - 78 no legal pair 13 17 2 72 0 0 - - 102 no legal pair 17 19 2 1088 0 0 - - 114 no legal pair 19 23 3 11784 234 62 96 138 138 EQUALITY ok 6q rule on run-LEGAL pairs at fold 23 (note: the raw qualifying-pair minimum is 96 < 6q = 138; alternation is what lifts it, exactly as kappa-not-L.md records at fold 29) 23 29 2 243816 288 0 120 - 174 no legal pair (note: the raw qualifying-pair minimum is 120 < 6q = 174; alternation is what lifts it, exactly as kappa-not-L.md records at fold 29) ============================================================================== PART 1 — the two exact transport SHAPES, and the SUM-shape identity ============================================================================== MAX-shape (copy theorem, cited): max_a Phi(T\{a,a-2}) = Phi(new). SUM-shape (tested here): sum_a Psi(T\{a,a-2}) = Psi(new) + bdy, for Psi any COUNT of windows. Both have index cost zero: the window size m does not grow. Below, Psi = D (slot count), Psi = total span, Psi = sum of g^2, and Psi = tail count N(theta) at three thresholds. x q Psi sum_a over deleted true new rel err 7 11 D (slot count) 135.000000000 135.000000000 0.000000 7 11 total span 2310.00000000 2310.00000000 0.000000 7 11 sum g^2 51732.0000000 51444.0000000 0.005598 7 11 N(theta>=60) 0.00000000000 0.00000000000 0 7 11 N(theta>=120) 0.00000000000 0.00000000000 0 7 11 N(theta>=180) 0.00000000000 0.00000000000 0 11 13 D (slot count) 1485.00000000 1485.00000000 0.000000 11 13 total span 30030.0000000 30030.0000000 0.000000 11 13 sum g^2 822276.000000 822276.000000 0.000000 11 13 N(theta>=60) 24.0000000000 24.0000000000 0.000000 11 13 N(theta>=120) 0.00000000000 0.00000000000 0 11 13 N(theta>=180) 0.00000000000 0.00000000000 0 13 17 D (slot count) 22275.0000000 22275.0000000 0.000000 13 17 total span 510510.000000 510510.000000 0.000000 13 17 sum g^2 16439364.0000 16439364.0000 0.000000 13 17 N(theta>=60) 874.000000000 874.000000000 0.000000 13 17 N(theta>=120) 0.00000000000 0.00000000000 0 13 17 N(theta>=180) 0.00000000000 0.00000000000 0 17 19 D (slot count) 378675.000000 378675.000000 0.000000 17 19 total span 9699690.00000 9699690.00000 0.000000 17 19 sum g^2 359985708.000 359985708.000 0.000000 17 19 N(theta>=60) 25246.0000000 25246.0000000 0.000000 17 19 N(theta>=120) 322.000000000 322.000000000 0.000000 17 19 N(theta>=180) 0.00000000000 0.00000000000 0 19 23 D (slot count) 7952175.00000 7952175.00000 0.000000 19 23 total span 223092870.000 223092870.000 0.000000 19 23 sum g^2 9271859796.00 9271859796.00 0.000000 19 23 N(theta>=60) 745674.000000 745674.000000 0.000000 19 23 N(theta>=120) 22790.0000000 22790.0000000 0.000000 19 23 N(theta>=180) 146.000000000 146.000000000 0.000000 ============================================================================== PART 2 — CANDIDATE TABLE, level by level ============================================================================== [1.8s] stats T_7 done [1.8s] stats T_11 done [1.8s] stats T_13 done [1.8s] stats T_17 done [1.8s] stats T_19 done [2.5s] stats T_23 done ------------------------------------------------------------------------------ 2.1 Sigma g^k : the polynomial moments ------------------------------------------------------------------------------ The bound they give is Phi_k = (Sigma g^k)^{1/k} >= G2. Overshoot is ln(Phi_k / G2), in nats; the LIFETIME budget is 0.598 to 1.19 nats (gate-multiplies.md section 5). k* is the least k with Phi_k <= 2*G2. x G2 mbar lnD Phi_2 ovr_2 Phi_4 ovr_4 Phi_8 ovr_8 k* k*/lnD 7 30 14.00 2.71 61 0.708 38 0.226 33 0.089 3 1.11 11 42 17.11 4.91 227 1.686 81 0.653 53 0.239 4 0.82 13 66 20.22 7.30 907 2.620 184 1.023 97 0.389 6 0.82 17 108 22.92 10.01 4055 3.625 434 1.392 178 0.499 7 0.70 19 150 25.61 12.84 18973 4.840 1027 1.924 302 0.700 9 0.70 23 204 28.05 15.89 96290 6.157 2468 2.493 501 0.898 10 0.63 per-fold behaviour of Sigma g^k (the multiplicative loss): fold q Sg2 ratio Sg4 ratio Sg8 ratio (q-2) ln(ratio)/k for k=2,4,8 budget 2lnq/q 11 13.874 21.240 48.873 9 1.3150 0.7640 0.4862 0.4360 13 15.984 26.788 123.482 11 1.3858 0.8220 0.6020 0.3946 17 19.993 31.305 124.384 15 1.4977 0.8609 0.6029 0.3333 19 21.898 31.207 69.121 17 1.5432 0.8602 0.5295 0.3099 23 25.756 33.400 56.834 21 1.6243 0.8771 0.5050 0.2727 ------------------------------------------------------------------------------ 2.2 Sigma exp(lambda g) : the exponential moment ------------------------------------------------------------------------------ Phi = ln Z / lambda >= G2 requires lambda large; at lambda with bite, Z is dominated by the record and the recursion becomes G2's own. x lambda*mbar Z Phi=lnZ/lam G2 ovr nats share of Z from the top gap 7 0.50 25.6403 90.8 30 1.108 0.227731 7 1.00 47.7386 54.1 30 0.590 0.357102 7 2.00 222.969 37.8 30 0.232 0.651700 7 4.00 11163.1 32.6 30 0.084 0.945738 11 0.50 231.830 186.4 42 1.490 0.058869 11 1.00 436.492 104.0 42 0.907 0.106679 11 2.00 2093.75 65.4 42 0.443 0.258897 11 4.00 120097 50.0 42 0.175 0.611662 13 0.50 2572.61 317.6 66 1.571 0.023852 13 1.00 5038.93 172.4 66 0.960 0.062268 13 2.00 32184.4 104.9 66 0.464 0.254907 13 4.00 8.12864e+6 80.4 66 0.198 0.690008 17 0.50 38973.7 484.5 108 1.501 0.005414 17 1.00 80734.9 259.0 108 0.875 0.027575 17 2.00 915545 157.3 108 0.376 0.270671 17 4.00 3.79487e+9 126.4 108 0.157 0.809128 19 0.50 668135 687.1 150 1.522 0.000559 19 1.00 1.44926e+6 363.4 150 0.885 0.004821 19 2.00 2.50582e+7 218.2 150 0.375 0.097392 19 4.00 5.80130e+11 173.5 150 0.145 0.513327 23 0.50 1.41155e+7 923.7 204 1.510 0.000011 23 1.00 3.16014e+7 484.5 204 0.865 0.000182 23 2.00 7.09099e+8 285.9 204 0.337 0.011679 23 4.00 6.67582e+13 223.3 204 0.090 0.256821 ------------------------------------------------------------------------------ 2.3 the gap-count vector by value class mod 6M ------------------------------------------------------------------------------ BITE, decided in one line and confirmed numerically: the class vector is unchanged by moving 6M from one gap to another, and that moves G2 freely. The numbers below show how close to uniform the classes already are. x M classes max/min class share chi^2/dof vs uniform does it see G2? 7 5 5 n/a 3.000 no 7 7 7 n/a 3.800 no 7 11 11 n/a 4.440 no 11 5 5 10.0000 14.667 no 11 7 7 14.0000 17.365 no 11 11 11 n/a 24.087 no 13 5 5 6.1250 107.848 no 13 7 7 8.4000 110.291 no 13 11 11 n/a 177.308 no 17 5 5 5.2469 1346.831 no 17 7 7 6.0790 1135.945 no 17 11 11 165.4500 2051.123 no 19 5 5 4.7192 20176.394 no 19 7 7 4.6991 14195.559 no 19 11 11 47.8199 28103.749 no 23 5 5 4.3569 385475.301 no 23 7 7 4.0590 237808.174 no 23 11 11 25.2855 495546.191 no ------------------------------------------------------------------------------ 2.4 the adjacent-large-pair (run-seed) count ------------------------------------------------------------------------------ S2(theta) = #{i : g_i + g_{i+1} >= theta}. The A5 Theorem A form is theta = 6q: a kill run of length 3 needs an adjacent qualifying pair, whose sum is at least 6q, so S2(6q) = 0 forces L <= 2. x q S2(6q) #adj qual pairs L S2(6q)=0 would force L<=2 min(g_i,g_{i+1}) max 7 11 0 0 1 YES (vacuous, L=1) 12 11 13 0 0 2 YES (vacuous, L=2) 30 13 17 0 0 2 YES (vacuous, L=2) 42 17 19 124 0 2 no 66 19 23 2192 234 3 no 78 23 29 6104 288 2 no 102 ------------------------------------------------------------------------------ 2.5 centered maxsum C_M = max_{k<=M} (maxsum_k - k*mbar) ------------------------------------------------------------------------------ Centering is what would cancel the fold's mean drift. The per-fold multiplier is the test: it must beat G2's own multiplier to be worth anything, since the bound it gives is G2 <= C_M + mbar. x mbar C_1 C_2 C_4 C_8 | ratio to previous level: G2 C_1 C_2 C_4 C_8 7 14.00 16.0 16.0 24.0 28.0 | 11 17.11 24.9 31.8 44.7 53.3 | 1.400 1.556 1.986 1.861 1.905 13 20.22 45.8 55.6 77.3 77.3 | 1.571 1.839 1.748 1.731 1.450 17 22.92 85.1 104.2 106.3 106.3 | 1.636 1.859 1.875 1.375 1.375 19 25.61 124.4 134.8 134.8 173.1 | 1.389 1.462 1.294 1.268 1.628 23 28.05 175.9 177.9 235.8 303.6 | 1.360 1.415 1.320 1.750 1.754 ------------------------------------------------------------------------------ 2.6 the alphabet: size, holes, and the dense prefix ------------------------------------------------------------------------------ A(T) = set of gap values. All are multiples of 6. If A had no holes below G2 then |A| = G2/6 and bounding |A| would bound G2. x G2 |A| G2/6 holes dense prefix (all 6k present up to) |A|/(G2/6) 7 30 4 5 1 18 0.8000 11 42 7 7 0 42 1.0000 13 66 10 11 1 48 0.9091 17 108 17 18 1 96 0.9444 19 150 23 25 2 108 0.9200 23 204 33 34 1 138 0.9706 ------------------------------------------------------------------------------ 2.7 the exponential moment IS the tail-count profile, in another basis ------------------------------------------------------------------------------ Z(lambda) = SUM_i e^{lambda g_i} = lambda * INT e^{lambda theta} N(theta) dtheta, so Z and N are a Laplace pair and their transports are the same inequality read in two bases. The index cost shows up as the ratio of the m-window partition function Z_m to Z_1: that ratio is what the merge term multiplies by, and it is the per-fold index price in nats. x lam*mbar Z_1 Z_2/Z_1 Z_3/Z_2 Z_4/Z_3 ln(Z_2/Z_1)/lam in mbar units 7 4.00 11163.1 36.619 520.915 54.387 0.900 11 4.00 120097 322.154 400.109 74.171 1.444 13 4.00 8.12864e+6 285.776 1357.177 49.032 1.414 17 4.00 3.79487e+9 510.637 60.203 66.692 1.559 19 4.00 5.80130e+11 126.055 62.812 30.133 1.209 23 4.00 6.67582e+13 83.873 1699.602 2073.112 1.107 ------------------------------------------------------------------------------ 2.8 the alphabet is an interval minus at most two values, and it only grows ------------------------------------------------------------------------------ Every old gap survives in at least q-4 > 0 copies, so A(new) CONTAINS A(old): PROVEN, and the dense prefix is non-decreasing. That transports with zero loss and it is a LOWER bound on G2, which is the wrong direction for the Zone Postulate. ok A(T_11) contains A(T_7) ok A(T_13) contains A(T_11) ok A(T_17) contains A(T_13) ok A(T_19) contains A(T_17) ok A(T_23) contains A(T_19) ============================================================================== PART 3 — IS ANY HISTOGRAM MARGINAL CLOSED UNDER FOLDING? ============================================================================== The A9 operator is exact but reads the gap WORD. Every candidate above except the maxsum family and the centered maxsum is a function of the HISTOGRAM alone. So: shuffle the old word, which fixes the histogram exactly and destroys nothing else about the multiset, then fold. If the new histogram moves, no histogram marginal can transport. x q seed new G2 new D new Sigma g^2 new N(>=120) new maxsum_2 7 11 true 42 135 51444.000 0 66 7 11 1 48 135 52884.000 0 72 7 11 2 42 135 51588.000 0 72 7 11 3 48 135 52452.000 0 66 7 11 4 48 135 52596.000 0 78 7 11 5 48 135 52740.000 0 84 11 13 true 66 1485 822276.00 0 96 11 13 1 72 1485 819612.00 0 138 11 13 2 78 1485 832068.00 0 102 11 13 3 66 1485 827244.00 0 126 11 13 4 72 1485 827532.00 0 102 11 13 5 102 1485 823572.00 0 144 13 17 true 108 22275 16439364 0 150 13 17 1 126 22275 16478820 1 216 13 17 2 144 22275 16441884 3 168 13 17 3 168 22275 16474500 7 228 13 17 4 156 22275 16436484 4 174 13 17 5 150 22275 16515180 7 180 17 19 true 150 378675 3.5998571e+8 322 186 17 19 1 186 378675 3.5998852e+8 270 258 17 19 2 222 378675 3.6016628e+8 308 264 17 19 3 204 378675 3.6017629e+8 310 252 17 19 4 192 378675 3.6030229e+8 325 252 17 19 5 222 378675 3.6041159e+8 311 264 19 23 true 204 7952175 9.2718598e+9 22790 234 19 23 1 282 7952175 9.2889958e+9 21292 336 19 23 2 324 7952175 9.2889657e+9 21095 366 19 23 3 276 7952175 9.2879247e+9 20993 354 19 23 4 330 7952175 9.2893051e+9 21109 360 19 23 5 324 7952175 9.2888790e+9 21171 372 ============================================================================== PART 4 — THE PROVEN TAIL-COUNT TRANSPORT, and what it certifies ============================================================================== N_new(theta) <= (q-2)*N(theta) + 2*SUM_{L>=1} Q_L(theta) Q_L(theta) = #{i : G_{L+1}(i) >= theta, g_{i+1}..g_{i+L-1} all qualify} PROVEN from A9 step 3/4 (nu_q(i,0) <= q-2, nu_q(i,L) <= 2 for L >= 1). Certified bound: G2(new) < min{theta : RHS(theta) < 1}. x q G2(old) certified G2(new) true G2(new) loss units loss nats maxsum_{L+1}(old) its loss nats L deciding 7 11 30 42 42 0 0.000 42 0.000 1 11 13 42 66 66 0 0.000 96 0.375 1 13 17 66 108 108 0 0.000 138 0.245 2 17 19 108 150 150 0 0.000 168 0.113 1,2 19 23 150 204 204 0 0.000 228 0.111 3 23 29 204 270 258 12 0.045 300 0.151 3 EQUIVALENT READING. The certificate IS the largest sum of consecutive OLD gaps whose interior gaps all qualify mod q. So the fold's new record is determined, to 0 or 12 units, by one window statistic of the old word, with no kill count and no residue-deleted tile to evaluate. AND IT CERTIFIES THE ZONE POSTULATE AT EVERY FOLD IN REACH: x q certificate q^2 margin ln(q^2/cert) nats true slack ln(q^2/G2) nats 7 11 42 121 1.058 1.058 11 13 66 169 0.940 0.940 13 17 108 289 0.984 0.984 17 19 150 361 0.878 0.878 19 23 204 529 0.953 0.953 23 29 270 841 1.136 1.182 ------------------------------------------------------------------------------ 4.1 CUSTODY: the inequality holds at EVERY theta, not only at the endpoint ------------------------------------------------------------------------------ x q thetas tested max ratio N_new(theta)/RHS(theta) any violation? ok fold 11: 7 thetas, max N_new/RHS = 1.0000, violations = 0 ok fold 13: 11 thetas, max N_new/RHS = 1.0000, violations = 0 ok fold 17: 18 thetas, max N_new/RHS = 0.8881, violations = 0 ok fold 19: 25 thetas, max N_new/RHS = 0.8975, violations = 0 ok fold 23: 34 thetas, max N_new/RHS = 0.9180, violations = 0 ------------------------------------------------------------------------------ 4.2 where the certificate comes from: L=0 alone cannot move it ------------------------------------------------------------------------------ x q old G2 cert from L=0 only cert with L>=1 true new G2 L=0 share of RHS at the certificate 7 11 30 30 42 42 0.0000 11 13 42 42 66 66 0.0000 13 17 66 66 108 108 0.0000 17 19 108 108 150 150 0.0000 19 23 150 150 204 204 0.0000 23 29 204 204 270 258 0.0000 ------------------------------------------------------------------------------ 4.3 the index the certificate actually spends, and A5's cap in this coordinate ------------------------------------------------------------------------------ K = number of kills inside the deciding window. A window of sum theta carrying K kills needs K-1 mutual spans that qualify; each is >= 2q-2, and by the Alternation Lemma each ADJACENT pair of them sums to >= 6q. So K <= 1 + theta/(3q) — which is A5 Theorem B's cap 0.18p verbatim, re-derived without ever mentioning a kill run. x q certificate K spent cap 1+theta/(2q-2) cap 1+theta/(3q) A5 Thm B's L bound G2/(3q) 7 11 42 1 3.10 2.27 1.27 11 13 66 1 3.75 2.69 1.69 13 17 108 2 4.38 3.12 2.12 17 19 150 2 5.17 3.63 2.63 19 23 204 3 5.64 3.96 2.96 23 29 270 3 5.82 4.10 2.97 ============================================================================== PART 4B — DOES THE TAIL-COUNT FAMILY CHAIN? the decisive test ============================================================================== To iterate the inequality one must carry the whole window family S_n(theta) = #{i : G_n(i) >= theta}, n = 1..M whose transport is, by the same alignment count, S_m'(theta) <= q*S_m(theta) + SUM_{K>=1} c(m,K) * S_{m+K}(theta), c(1,K) = 2 (the kills of ONE new gap form a run, |A_L| <= 2) c(m,K) = 2(m+K-1) (m >= 2: the kills need not be consecutive) The chain below starts from the EXACT profiles of T_11 and carries M of them, dropping every term with m+K > M. Dropping is OPTIMISTIC, so the chain below is a LOWER bound on any honest certificate. M = 4 maxsum_M(T_11) = 108 (the ceiling the chain can never pass) fold q 13 17 19 23 29 chain 108 108 108 108 108 truth 66 108 150 204 258 q^2 169 289 361 529 841 verdict ok ok FALSE FALSE FALSE M = 8 maxsum_M(T_11) = 180 (the ceiling the chain can never pass) fold q 13 17 19 23 29 chain 180 180 180 180 180 truth 66 108 150 204 258 q^2 169 289 361 529 841 verdict BUSTS ok ok FALSE FALSE M = 12 maxsum_M(T_11) = 240 (the ceiling the chain can never pass) fold q 13 17 19 23 29 chain 240 240 240 240 240 truth 66 108 150 204 258 q^2 169 289 361 529 841 verdict BUSTS ok ok ok FALSE M = 16 maxsum_M(T_11) = 330 (the ceiling the chain can never pass) fold q 13 17 19 23 29 chain 330 330 330 330 330 truth 66 108 150 204 258 q^2 169 289 361 529 841 verdict BUSTS BUSTS ok ok ok READING: the chain's certificate is capped by maxsum_M(T_11) for every M, because S_M has no term feeding it and its support never grows. So a fixed-M truncation certifies a CONSTANT while the truth diverges, and it goes FALSE as soon as the truth passes that constant. The index M must grow with the level, at the rate 4.3 prices, which is the accumulating index of gate-multiplies.md sections 2 and 5. ============================================================================== PART 5 — THE PER-FOLD LOSS LEDGER, in nats, against the sharp budget ============================================================================== budget per fold = 2 ln q / q (gate-multiplies.md section 7); total lifetime slack 0.598 to 1.19 nats. A candidate whose per-fold loss does not sum to under a nat over the whole ladder is dead on arrival. fold q budget truth ln(G2'/G2) Sg2 Sg8 Z(4/mbar) C_1 tail-count transport 11 0.4360 0.3365 0.979 0.150 0.091 0.105 0.000 13 0.3946 0.4520 0.934 0.150 0.023 0.157 0.000 17 0.3333 0.4925 1.005 0.110 -0.041 0.127 0.000 19 0.3099 0.3285 1.215 0.201 -0.012 0.051 0.000 23 0.2727 0.3075 1.317 0.198 -0.055 0.039 0.000 cumulative excess over the truth, this ladder only: Sigma g^2 5.449 Sigma g^8 0.809 Z(4/mbar) 0.006 C_1 0.481 ============================================================================== PART 6 — T_29 AND THE DEEP END (streamed, nothing stored) ============================================================================== [4.4s] streaming T_23 -> T_29 D(T_29) = 214708725 (prod (p-2) = 214708725) G2(T_29) = 258 (recorded 258) mbar = 30.1324 Sigma g^2 = 2.936514993e+11 Phi_2 = 541896.2 overshoot 7.650 nats Sigma g^4 = 1.428032336e+15 Phi_4 = 6147.3 overshoot 3.171 nats Z(4/mbar) = 5.4128210e+15 Phi = 272.9 overshoot 0.056 nats |A| = 41 G2/6 = 43 holes = 2 dense prefix = 240 ok D(T_29) streamed ok G2(T_29) = 258 streamed ============================================================================== SELF-TEST SUMMARY ============================================================================== all checks passed [14.4s] done node research/attack-foldL-03-transport.js 23 14.34s user 0.15s system 100% cpu 14.485 total