Ancillary files for "A New Upper Bound for the First Autocorrelation Inequality and 306 Circle-Packing Records via
Active-Set Escape from Equioscillating and KKT Optima" (Lindell S. Cumes, QWNTL Labs, 2026).
Project page: https://qwntl.com/paper-new-constructions

All construction files are plain text; lines starting with '#' are comments; every decimal is an exact rational.
Every file ships under the name it was produced with, except that in three names (erdos_solo_e3.txt,
ac3_solo_c3_97k_5.txt and ac3-solo-final.txt) the single autocorrelation agent's label is replaced by 'solo'.
Compressed files have .xz appended, so decompressing one gives the name without .xz; the two C1 archive files are named
by their sha256 and have no extension. In the names, SX marks the shared archive and A<k> its attempt k; solo marks the
single autocorrelation agent.

erdos_solo_e3.txt   Erdos re-grid construction (2,560 cells). 2,560 heights in [0,1] with sum exactly 1280 = n/2:
                    a step function h on [0,2] (cell width 2/n), integral 1. M <= (2/n) max_k sum_i h_i (1 - h_{i+k})
                    = 487496381280743038886397039070859359383/(128*10^37) = 0.38085654787558049912999768677...
erdos-SX-A51.txt    Erdos, shared archive, attempt 51 (3,584 cells). 3,584 heights a_i in [0,1] with sum exactly
                    1792 = n/2, same format.
                    M <= 97499325154129259661263000479811180603/(256*10^36) = 0.38085673888331742055180859562...
49c83eb6f9f82e244578fab15d056757886679700bd8c8c653f0b100ae882f2d.xz
                    C1, shared archive, 40.75 h. 262,144 nonnegative heights on [-1/4,1/4]: the shared archive's best
                    C1 construction at 40.75 working hours (archive file, named by its sha256).
                    C1 <= 2n max_k c_k / S^2, c_k = sum_{i+j=k} a_i a_j, S = sum a_i; = 1.50266484077609912906120...
475b6c7645aeecf3b0b7248817ed67952c28c0e67145c9c546f9a3f270ee107d.xz
                    C1, single agent, 44.0 h. 262,144 nonnegative heights, same format: a single agent's best C1
                    construction at 44.0 working hours (archive file, named by its sha256);
                    = 1.50267149081134106921496...
ac1-SX-A43.txt.xz   C1, shared archive, 3.97 h. 262,144 nonnegative heights on [-1/4,1/4], same format: the
                    construction at 3.97 working hours; = 1.50274043592276697640985...
ac3_solo_c3_97k_5.txt.xz
                    C3+, 97,000 cells. 97,000 signed heights on [-1/4,1/4]. Signed constant: 2n max_k c_k / S^2
                    = 1.44886703134291290097... Within the Einstein Arena's 100,000-value input limit.
ac3-solo-final.txt.xz
                    C3+, 102,400 cells. 102,400 signed heights on [-1/4,1/4]. Signed constant: 2n max_k c_k / S^2
                    = 1.44888247044528093575... More than 100,000 values, so not an Einstein Arena-eligible input. Both
                    C3 files derive from MILO's published coarse vectors. Their absolute-value scores max|f*f|/(int f)^2
                    are 5.97 and 5.83, so they say nothing about that variant.
minratio14_3d.txt   14 points 'x y z'; (d_max/d_min)^2 = 4.16578347458007331688... (a polish of FICO's configuration).
                    Not discussed in the paper; verify_all.py checks it with audit_geo/exact.py.
csqv.tar.xz         csqv/csqvN.txt for the 306 sizes: N lines 'x y r', unit square [0,1]^2 in the corner frame (subtract
                    1/2 for Packomania's centred frame). Feasible as written: r > 0, r <= x, y <= 1 - r and
                    (xi-xj)^2 + (yi-yj)^2 >= (ri+rj)^2 for every pair, all exactly.
csqv_values.txt     N, exact sum of radii, previous best (located 2026-10-09 and still the best published value on
                    2026-10-30), margin, run.
verifier/           verify_all.py runs the checkers below on these files and compares with the paper's values:
                      cd anc && python3 verifier/verify_all.py          (everything; about 30 seconds on one core)
                      cd anc && python3 verifier/verify_all.py --quick  (analytic files and three packings)
                    Needs Python >= 3.9 and libgmp. The checkers are copied unchanged, except that the stand-alone
                    entry points of audits_crossfamily/exact.py, audit_geo/exact.py and audit_geo2/exact_all.py have
                    their default input paths changed; verify_all.py does not use those entry points, and SHA256SUMS
                    gives each file's hash as shipped and, for those three, as written. The first four are
                    independent exact checkers (written from the problem definitions; they share no code with
                    the search or the evaluator):
                      audits_crossfamily/exact.py  Erdos and C1: exact GMP convolution of carry-free packed integers,
                                                   plus direct integer sums
                      audit_geo/exact.py           decimal parser, all-pairs check, 14-point ratio
                      audit_geo2/exact_all.py      integer grid check of packings with a completeness proof
                      audit_c3/exact.py            signed and absolute C3 scores
                      builder/verify_indep.py      the packing agent's own verifier (packings)
                      builder/verify_minratio.py   the packing agent's own verifier (14 points)
                    verify_all.py itself was written by the paper's authors.
SHA256SUMS          sha256 of every file above as shipped (uncompressed). Construction files are byte-identical to the
                    files the searches wrote, except that the '#' comment line of 29 packing files (from the N+-1
                    transfer step) has a machine name removed from its "source=" field, and the '#' comment line of
                    ac3-solo-final.txt has the agent's label replaced by 'solo'. Their data lines are unchanged, and
                    for each of these 30 files SHA256SUMS also gives the sha256 of the file as the search wrote it.