# The exact grid zeros at K = 36 and K = 72 (eta = 1/36, alpha = 13/36): the runs of rigfam.py and indep_export.py, the K = 36 capture, and the separately written checks # Computer-assisted exact certificates of two finite grid problems behind section 8, with their separately written checks; not part of Theorem A's record. # Each section below is one output file. unpack.py writes it back out and checks its SHA-256: the original one where the # shipped text is the file as written by its run, and the shipped one where declared substitutions were made. ==== file: rigv4_K36_13.out | original SHA-256 6f5f69c83174363020fb08ca507ce7971ff564fb273669bdd6dbe4b5259112a8 | shipped SHA-256 6f5f69c83174363020fb08ca507ce7971ff564fb273669bdd6dbe4b5259112a8 | trailing newline: yes | as written by the run rigfam v4 verify 36 1 18 13 4 3 full; sources {'rigfam': '1a212ad4192861e1142ba232b1d04b180e32c306c819baa365201f1c928ee545', 'famrk': '0a34eca9a07d00498e2b7f5539a92f549adb61529f225446fa36a555585c2126', 'modm_otf_lowmem2': '2b6c25c17ca07b6155414e40092bb9d9ad651d7bdcb46a2bfa2afded6a8c33e6'}; python 3.14.4, numpy 2.5.2, scipy 1.18.0; float info: eps 2.220446049250313e-16, smallest subnormal 5e-324; hints: FAKE=None BASIS=None TEST hooks: [] exact distributions: Q 139 bits, 50058 omega entries, 7559 classes (j <= 36); prime class mass exactly 1/K; 0 s; RSS 0.11 GB family LP: 1899 rows (1527 Type I keys, 372 twisted), 7559 classes, nnz 1205622; max entry count 23; 0 s; RSS 0.18 GB float fake: C- = 7.461e-14, residual 1.04e-12 (row-equilibrated); 5 s exact residual of the dyadic point: max |r| 3.150e-14; 5 s; RSS 0.19 GB certify_core: ||I - C B||_inf <= 6.972e-09; |delta_S - x~|_inf <= eps = 7.497e-21 (exact rational); max |x~| 1.635e-11; exact minimum basis slack 5.039869e-03; prime exactly -1 and every off-basis class within [-1, 3] (checked exactly) VERDICT: VERIFIED EXACT FAKE; 6 s; RSS 0.20 GB certificate written: rigfam_cert_K36_m01_thI18_thT13_cap4_r3_full_1a212ad4.npz (sha256 7efe044f74914a8e0884d3852e5447edb0c0a3876c158d29cdbdf923bb2e9e26) RESULT rigfam v4 K=36 m0=1 thI=18 thT=13 cap=4 ct+s3 full family: VERIFIED EXACT FAKE ==== end: rigv4_K36_13.out ==== file: v4_capture_K36.log | original SHA-256 0f4ed730e1e572983492ebd0e866464468db760360122af6f58f97b733f779fe | shipped SHA-256 a294d12689ac173e03ec0498aee34cadcec365762150a91ff4d2689b1d627f5a | trailing newline: yes | 2 declared substitution(s): the directory paths of the two input files removed REVIEW CONTROLS: Python 3.14.4, NumPy 2.5.2, SciPy 1.18.0; nice=10; affinity=[0]; AS limit=2900000000 bytes; BLAS/OpenMP/HiGHS threads=1; bytecode disabled exact distributions: Q 139 bits, 50058 omega entries, 7559 classes (j <= 36); prime class mass exactly 1/K; 0 s; RSS 0.10 GB family LP: 1899 rows (1527 Type I keys, 372 twisted), 7559 classes, nnz 1205622; max entry count 23; 1 s; RSS 0.18 GB float fake loaded from rigfam_cert_K36_m01_thI18_thT13_cap4_r3_full_1a212ad4.npz (untrusted) float fake: C- = 0.000e+00, residual 1.04e-12 (row-equilibrated); 1 s exact residual of the dyadic point: max |r| 3.150e-14; 1 s; RSS 0.18 GB basis hint: review_v4/K36_basis_hint.npy (1899 indices) certify_core: ||I - C B||_inf <= 6.972e-09; |delta_S - x~|_inf <= eps = 7.497e-21 (exact rational); max |x~| 1.635e-11; exact minimum basis slack 5.039869e-03; prime exactly -1 and every off-basis class within [-1, 3] (checked exactly) VERDICT: VERIFIED EXACT FAKE; 2 s; RSS 0.19 GB Captured K36 basis enclosures and inverse; recomputed result equals saved certificate REVIEW peak RSS=251416 KiB ==== end: v4_capture_K36.log ==== file: v4_exact_grid_K36.log | original SHA-256 5b248dfca17625c015268c093b0d0993ac9fdf9cb281398065852cc0862d78e0 | shipped SHA-256 5b248dfca17625c015268c093b0d0993ac9fdf9cb281398065852cc0862d78e0 | trailing newline: yes | as written by the run REVIEW CONTROLS: Python 3.14.4, NumPy 2.5.2, SciPy 1.18.0; nice=10; affinity=[0]; AS limit=2900000000 bytes; BLAS/OpenMP/HiGHS threads=1; bytecode disabled independent K36: independently enumerated through n=18; 385 shapes; 0.0s independent K36: independently enumerated through n=24; 6126 shapes; 0.0s independent K36: independently enumerated through n=30; 27417 shapes; 0.1s independent K36: independently enumerated through n=36; 97921 shapes; 0.3s independent K36: ALL exact grid and coefficient/residual enclosure checks PASS {'shapes_enumerated': 97921, 'shapes_at_K': 17977, 'feature_entries': 50058, 'rows': 1899, 'classes': 7559, 'original_rows': {'I': 1597, 'T': 373}, 'basis_positions_checked': 109541, 'seconds': 0.6769316860008985} REVIEW peak RSS=94684 KiB ==== end: v4_exact_grid_K36.log ==== file: v4_independent_K36_bound.log | original SHA-256 2c97a96617397b925a8c99ffaaf57a1da537805a6ea5de683b818a255e2f3e85 | shipped SHA-256 2c97a96617397b925a8c99ffaaf57a1da537805a6ea5de683b818a255e2f3e85 | trailing newline: yes | as written by the run REVIEW CONTROLS: Python 3.14.4, NumPy 2.5.2, SciPy 1.18.0; nice=10; affinity=[0]; AS limit=2900000000 bytes; BLAS/OpenMP/HiGHS threads=1; bytecode disabled K36 VERIFIED BY INDEPENDENT OUTWARD BOUNDS nE <= 6.9650296499014855e-09 (stored 6.972344046861355e-09); zmax <= 7.4515543047423815e-21 (stored 7.4974590579697905e-21); eps <= 7.4515543566426787e-21; exact minslack=0.0050398689100726417 Every upper-bound scalar operation rounds outward; exact Fraction endpoints passed for every class; dot products use the standard IEEE error bound with exact gamma and explicit absolute underflow allowance. REVIEW peak RSS=153412 KiB ==== end: v4_independent_K36_bound.log ==== file: K36_independent_bounds.json | original SHA-256 32fb611c78e0c2f0887e52c6252287d8f9b7d60c26c17262390f180f088f0cf6 | shipped SHA-256 32fb611c78e0c2f0887e52c6252287d8f9b7d60c26c17262390f180f088f0cf6 | trailing newline: yes | as written by the run { "nE": 6.9650296499014855e-09, "zmax": 7.451554304742381e-21, "eps": "4952407296987589/664613993263401735391370636167217152", "minslack": "7542555251406260511304292422541870336895896126305/1496577666203136697153923392708496565296305794973696", "stored_nE": 6.972344046861355e-09, "stored_zmax": 7.49745905796979e-21, "stored_eps": "1245729059638081/166153498314635121196424830128226304", "verdict": "K36 VERIFIED BY INDEPENDENT OUTWARD BOUNDS" } ==== end: K36_independent_bounds.json ==== file: rigv4_K72_26.out | original SHA-256 470140277d0b8691c72f090705994ad1a45ef051897424d51d924add50381b1e | shipped SHA-256 470140277d0b8691c72f090705994ad1a45ef051897424d51d924add50381b1e | trailing newline: yes | as written by the run rigfam v4 verify 72 2 36 26 4 3 full; sources {'rigfam': '1a212ad4192861e1142ba232b1d04b180e32c306c819baa365201f1c928ee545', 'famrk': '0a34eca9a07d00498e2b7f5539a92f549adb61529f225446fa36a555585c2126', 'modm_otf_lowmem2': '2b6c25c17ca07b6155414e40092bb9d9ad651d7bdcb46a2bfa2afded6a8c33e6'}; python 3.14.4, numpy 2.5.2, scipy 1.18.0; float info: eps 2.220446049250313e-16, smallest subnormal 5e-324; hints: FAKE=../rerun/famrk_K72_m02_thI36_thT26_cap4_ct+s3.npz BASIS=../rigv2/S_K72_26_s3_hint.npy TEST hooks: [] exact distributions: Q 345 bits, 743270 omega entries, 68615 classes (j <= 72); prime class mass exactly 1/K; 4 s; RSS 0.47 GB family LP: 15282 rows (13003 Type I keys, 2279 twisted), 68615 classes, nnz 33744702; max entry count 123; 15 s; RSS 2.12 GB float fake loaded from ../rerun/famrk_K72_m02_thI36_thT26_cap4_ct+s3.npz (untrusted) float fake: C- = 1.710e-13, residual 3.79e-12 (row-equilibrated); 15 s exact residual of the dyadic point: max |r| 6.251e-14; 24 s; RSS 1.94 GB basis hint: ../rigv2/S_K72_26_s3_hint.npy (15282 indices) certify_core: ||I - C B||_inf <= 5.525e-04; |delta_S - x~|_inf <= eps = 2.615e-13 (exact rational); max |x~| 1.342e-08; exact minimum basis slack 3.173078e-04; prime exactly -1 and every off-basis class within [-1, 3] (checked exactly) VERDICT: VERIFIED EXACT FAKE; 318 s; RSS 1.47 GB certificate written: rigfam_cert_K72_m02_thI36_thT26_cap4_r3_full_1a212ad4.npz (sha256 68b3da5fe80cb97d7f4e7e2e57f195a33f2d10827b0d55a1fdc2f14ed941b1a8) RESULT rigfam v4 K=72 m0=2 thI=36 thT=26 cap=4 ct+s3 full family: VERIFIED EXACT FAKE maxRSS 3857532 KB, user 278.39 s ==== end: rigv4_K72_26.out ==== file: indep_export_K72.out | original SHA-256 bf938acbad43327d1b78ac495a45cd59d5c7e4a5d9f2b775c351c792cc747719 | shipped SHA-256 bf938acbad43327d1b78ac495a45cd59d5c7e4a5d9f2b775c351c792cc747719 | trailing newline: yes | as written by the run exact distributions: Q 345 bits, 743270 omega entries, 68615 classes (j <= 72); prime class mass exactly 1/K; 3 s; RSS 0.47 GB family LP: 15282 rows (13003 Type I keys, 2279 twisted), 68615 classes, nnz 33744702; max entry count 123; 12 s; RSS 2.00 GB float fake loaded from ../rerun/famrk_K72_m02_thI36_thT26_cap4_ct+s3.npz (untrusted) float fake: C- = 1.710e-13, residual 3.79e-12 (row-equilibrated); 12 s exact residual of the dyadic point: max |r| 6.251e-14; 22 s; RSS 1.82 GB basis hint: ../rigv2/S_K72_26_s3_hint.npy (15282 indices) certify_core: ||I - C B||_inf <= 5.525e-04; |delta_S - x~|_inf <= eps = 2.615e-13 (exact rational); max |x~| 1.342e-08; exact minimum basis slack 3.173078e-04; prime exactly -1 and every off-basis class within [-1, 3] (checked exactly) VERDICT: VERIFIED EXACT FAKE; 613 s; RSS 1.48 GB indep_export: verdict VERIFIED EXACT FAKE; files in K72_indep; claims 0.0005524641287095424 2.613395204845305e-13 2.614839810042675e-13 maxRSS 3860764 KB, user 538.54 s ==== end: indep_export_K72.out ==== file: K72_indep_README.md | original SHA-256 3757799cab44a52208b5b23c8091df45b7181f7b473bdecf521b4cd9c0f6dfd6 | shipped SHA-256 a62960c6749e4fa709f180dd3b9373e5a96f20092332bb6f8904cd6a45b1ab93 | trailing newline: yes | 1 declared substitution(s): an author name removed from the title line # K72_indep: inputs for an independent check of the K = 72 rigfam certificate (Wed 11 Nov 2026) Produced by indep_export.py (sha256 d85fbc47..., in the parent directory). It ran the UNMODIFIED rigfam v4 (sha256 1a212ad4...) `verify` for (K, m0, thI, thT, cap, r, J) = (72, 2, 36, 26, 4, 3, full), with FAKE = rk/rerun/famrk_K72_m02_thI36_thT26_cap4_ct+s3.npz (float starting point, untrusted) and BASIS = rk/rigv2/S_K72_26_s3_hint.npy (hint, untrusted). It saved the core's inputs exactly as the core received them. C was recomputed by the same LAPACK calls on the same input: dgetrf then dgetri on B.toarray(order="F"), one thread. The core's own x~ is BITWISE equal to the exported xt = C @ (-r), which the script asserts (claims.json: core_xt_bitwise_equal_to_exported_xt = true). Verdict of that run: VERIFIED EXACT FAKE (log ../indep_export_K72.out). ## Arrays (all .npy; binary64 = IEEE double) | name | shape | dtype | meaning | |---|---|---|---| | B_data, B_indices, B_indptr | (1576545,), (1576545,), (15283,) | float64, int32, int32 | CSR of B, shape (m, m), m = 15282: the float basis matrix the core used (rows scaled by p2, columns S) | | dB_data, dB_indices, dB_indptr | same | float64, int32, int32 | CSR of dB >= 0, the entrywise radius of B (same sparsity pattern as B) | | r, dr | (15282,) | float64 | the float residual enclosure: the exact residual r_ex (scaled by p2) satisfies \|r_ex - r\| <= dr | | C | (15282, 15282) | float64, C order | the float approximate inverse of B (about 1.87 GB) | | xt | (15282,) | float64 | the float correction x~ = C(-r) | | S | (15282,) | int64 | basis class indices (into the class list), sorted; the prime class cP = 1 is not among them | | g | (68615,) | float64 | the dyadic point (multiples of 2^-60), g[cP] = -1; all classes | | p2 | (15282,) | float64 | exact powers of two, the row scaling: B = diag(p2) A~[:, S], r = p2 * r~ | | class_codes | (68615,) | int64 | the sorted class codes (famrk.encoder(72, 2, 26, 3) of (j, tau, k, rmax, L, top1, top2, top3)) | | row_keys | (15282, 10) | int64 | rows in order: (type 0 = Type I / 1 = twisted, n = K - \|v\|, j, tau, k, rmax, L, top1, top2, top3) | | Rex_hex.txt | 15282 lines | hex ints | exact integers R_i = Q 2^SH (A_ex g)_i, unscaled; Q and SH are in claims.json | claims.json: cP = 1, cap = 4, m, nC, nE = 5.524641287095424e-4, zmax = 2.613395204845305e-13, eps (an exact fraction, eps_fraction = zmax/(1 - nE) as Fractions of the two floats), Q (hex), SH = 60, the source shas. ## The inequalities, in order (C, B, dB, r, dr, xt, g are exact dyadic rationals as stored) 0. Premises established upstream by `verify` (checkable independently from the exact data): - (P1) |B_ex - B| <= dB entrywise. Here B_ex[i, k] = p2[i] A_ex[i, S[k]], and A_ex[i, c] is the exact family-matrix entry: the sum, over the omega with merge(row_keys[i], omega) = class c, of D_ex(omega)/Q for a Type I row, or Dp_ex(omega)/Q for a twisted row. D and Dp are the exact distributions of rigfam.DistExact. - (P2) |r_ex - r| <= dr, where r_ex[i] = p2[i] R_i / (Q 2^SH). 1. Contraction. N := max_i [ sum_k |(I - C B)[i, k]| + sum_k (|C| dB)[i, k] ] <= nE < 1, in the infinity norm (maximum absolute row sum). Then ||I - C B_ex||_inf <= N < 1 for every B_ex allowed by (P1), so B_ex is invertible. 2. Residual of x~. With rho := -r - B xt and rho_bar := |rho| + dr + dB |xt| (entrywise), check Z := max_i (|C| rho_bar)[i] <= zmax. Then ||C(-r_ex - B_ex xt)||_inf <= Z for every r_ex, B_ex allowed by (P1) and (P2). 3. Correction. eps = zmax/(1 - nE), exact. The unique delta_S with B_ex delta_S = -r_ex satisfies |delta_S - xt|_inf <= eps (Rump / Krawczyk: e = delta_S - xt solves e = C(-r_ex - B_ex xt) + (I - C B_ex) e). 4. Bounds, exact. For every k: g[S[k]] + xt[k] - eps >= -1 and g[S[k]] + xt[k] + eps <= cap - 1 = 3. For every class c not in S with c != cP: -1 <= g[c] <= 3. Also g[cP] = -1 and cP is not in S. 5. Conclusion. g* = g + delta (delta = delta_S on S, 0 elsewhere) satisfies A_ex g* = 0 on every key row, since diag(p2) is invertible and A_ex g* = A_ex g + A_ex[:, S] delta_S = (r_ex - r_ex)/p2. It also satisfies g*(prime) = -1 and the bounds. Every free-LP row is a positive multiple of its key row (twisted rows up to sign), so the free K = 72 shape LP at (1/36, 26/72), caps 4, theta = K/2, has C- = 0 exactly. ## Rounding convention of the claims binary64, round-to-nearest-even, subnormals on. The claimed nE and zmax were computed by the core (rigfam v4, _certify_core_checked) in round-to-nearest with a priori gamma_n inflations and an absolute underflow allowance; they are claims to be re-checked. An independent check of steps 1-2 should use exact or outward-rounded arithmetic (e.g. split |C| and B into exactly representable parts, or interval matrix products with directed rounding). Step 4 is exact (Fractions); step 3 is one exact division. ==== end: K72_indep_README.md ==== file: K72_indep_claims.json | original SHA-256 6b6835be68ddff20ac5c99ceefb505547762547f8dc02bb815363d14393b6fa6 | shipped SHA-256 6b6835be68ddff20ac5c99ceefb505547762547f8dc02bb815363d14393b6fa6 | trailing newline: no | as written by the run { "cP": 1, "cap": 4, "m": 15282, "nC": 68615, "B_format": "CSR (data, indices, indptr), shape (m, m)", "B_is_csr": true, "dB_is_csr": true, "core_verdict": "VERIFIED EXACT FAKE", "core_xt_bitwise_equal_to_exported_xt": true, "getrf_info": 0, "getri_info": 0, "nE": 0.0005524641287095424, "zmax": 2.613395204845305e-13, "eps_fraction": "2588181250043537/9898048974561454564885659648", "eps_float": 2.614839810042675e-13, "minslack_basis": "28289166020688429615385128574264737106973/89153699347059773639751249517707257305890816", "verdict": "VERIFIED EXACT FAKE", "params": { "K": 72, "m0": 2, "thI": 36, "thT": 26, "cap": 4, "r": 3, "J": "full" }, "Qhex": "0x1b5705796695b6477f3757a3d607aa1ae7355aa9cbff75b861f20f84a9922d720c518400000000000000000", "SH": 60, "sources": { "rigfam": "1a212ad4192861e1142ba232b1d04b180e32c306c819baa365201f1c928ee545", "famrk": "0a34eca9a07d00498e2b7f5539a92f549adb61529f225446fa36a555585c2126", "modm_otf_lowmem2": "2b6c25c17ca07b6155414e40092bb9d9ad651d7bdcb46a2bfa2afded6a8c33e6", "indep_export": "d85fbc4767ef32626964fdef6b685252b4f1565326528f5c163fc5095cac0923" }, "FAKE": "../rerun/famrk_K72_m02_thI36_thT26_cap4_ct+s3.npz", "BASIS": "../rigv2/S_K72_26_s3_hint.npy" } ==== end: K72_indep_claims.json ==== file: K72_indep_READY.txt | original SHA-256 ecbca8b47468786a907ad20768f55e589f61e7c1d1caf9ca82f8a04a5d64257a | shipped SHA-256 ecbca8b47468786a907ad20768f55e589f61e7c1d1caf9ca82f8a04a5d64257a | trailing newline: yes | as written by the run 4f30b260c9f2f9fa7021621809c2b77b52e751f9efe7793f766244ff7d805252 B_data.npy e9e7bbaf6573ecc9c4c4c48814445df608e117cfeeb06f565db8bfac90361f4b B_indices.npy 69194d483e131935f24341ed988730148e6617430eb820ceb758ccce76ac615e B_indptr.npy 907b2981e2c81202388a5b273a3190a1b19d2bbe0e21ed8b2aae7e0e742cfb0c C.npy 38a1c6c5130014a9b640a00e2042452d303a91a65ca12f61dd39094ab233eace S.npy 7a10f7b3e6cde903f1b353b745f78fa506d7bbf6bef38004ae3e40b58b811a2b class_codes.npy 1e3617fe0758eac7bd8dc9104d42cb4ebd188cee77daace3830b2ee02ba76a9d dB_data.npy e9e7bbaf6573ecc9c4c4c48814445df608e117cfeeb06f565db8bfac90361f4b dB_indices.npy 69194d483e131935f24341ed988730148e6617430eb820ceb758ccce76ac615e dB_indptr.npy 01ce0c7d225f8297547645f1521b6594350e163a301090f969466e4a078e14c6 dr.npy 890304d0022ad31ddc6b6240c9e6331dd0d16d9c17059555f6f121acb10f0624 g.npy 073e78c072ed8842e4cfdeb122da83e9e418c593daff197f2dab81c3867981b0 p2.npy 9114d419489898779a81bee8e588f28892580f65d6f4b1e602bdd34cc4a18158 r.npy 13b71accec2507aa39daa09ca3b5c0e6e39bbbd33afbc4cbbccda1945b67a09b row_keys.npy a5b4f240f8cf0ebf590fcfb80452014d4f1240d409f4fa480b4e0c85244e97b9 xt.npy 4cfbf116b7a845cb15afc32a3f2881ab53215aee20f5fc07ac84ff3dbf153bb1 Rex_hex.txt 6b6835be68ddff20ac5c99ceefb505547762547f8dc02bb815363d14393b6fa6 claims.json 3757799cab44a52208b5b23c8091df45b7181f7b473bdecf521b4cd9c0f6dfd6 README_INDEP.md ==== end: K72_indep_READY.txt ==== file: v4_independent_K72.log | original SHA-256 c288118231fa38037d5f85c8ac18e9890f622dc7343927359115d514e60e437f | shipped SHA-256 c288118231fa38037d5f85c8ac18e9890f622dc7343927359115d514e60e437f | trailing newline: yes | as written by the run total used free shared buff/cache available Mem: 31 8 2 0 20 22 Swap: 32 1 30 PART C ONLY: nice=10, affinity=[0], threads=1, RLIMIT_AS=8000000000; no simultaneous review process [0.9s] Every frozen READY hash independently rechecked before use [0.9s] Certificate fields: {'params': ((7,), 'int64'), 'Qhex': ((), '