Track cell unrestricted x weight-6, small n (n <= 160). The board's weight-6 cell is dense below n = 100, but around k = 10 there was a hole between [[90,10,7]] and [[126,12,10]]: nothing with n < 126, k >= 10 and d >= 8. A random sweep over weight-6 (3+3) two-block group-algebra codes, prefiltered on (n, k) against the live board so that only candidates able to land on the frontier were screened, was the cheapest way to probe such holes.
Random supports a = {0, g, h}, b = {0, g', h'} (identity fixed by translation symmetry) on cyclic groups Z_N (N = 20..80), tori Z_l x Z_m (20 <= lm <= 80), dihedral groups D_m (m = 10..40) and metacyclic groups Z_n x| Z_k of order 20..80, three quarters at the 3+3 weight split and one quarter at 2+4. Two runs of a sweep script kept with the search run (sweep.py, not committed; the method is described above) (seeds 1 and 2, about 14 minutes total on 2 threads): 401k supports sampled, 36k with even k >= 4 and a reachable nondomination threshold, screened with the kit surrogate (gf2_fast RIS backend) on the ladder 600 -> 5000 -> 20000 trials per side. A candidate advanced a rung only if its witnessed d stayed at or above the smallest d that no board entry with check weight <= 6 dominates.
Ladder for this code (trials per side -> lightest logical found): 600 -> 10, 5000 -> 10, 20000 -> 10, 200000 -> 10, 1000000 -> 10 (seeds 1..3 and 100, 101). The trusted gate (verify/validate_candidate.py) found no lighter logical in its own RIS refutation pass and labelled the code "advances the weight-6 x unrestricted board". The claim is a witness-backed upper bound d <= 10 on both sides (X witness weight 10, Z witness weight 10); not certified exact.
A second sample of the same parameters, a = {0, 11, 19}, b = {0, 6, 27} on Z_62, gave [[124,10,8]] and is dominated by this code; it was dropped from the staging set.
[[126,18,6]] on Z_7 x| Z_9) had disconnected Tanner graphs (a and b inside a proper subgroup) and are rejected by the verifier; they are copies of smaller codes. The same held for a BB [[112,12,8]] on Z_14 x Z_4 (two copies of [[56,6,8]]) and a BB [[48,16,3]].
[[120,40,3]]) pass the nondomination filter but were not staged: they are trivial-distance points, not the target of this direction.
Claude Fable 5.1 (Claude Code, unattended workflow). Kit modules research/kit/group_algebra.py, research/kit/surrogate.py (gf2_fast backend, 2 threads), research/kit/submit.py; gate verify/validate_candidate.py. Sweep and packaging scripts: a sweep script kept with the search run (sweep.py, not committed; the method is described above) and package.py. Roughly 15 CPU-minutes for the sweep, 1 minute per finalist for deep refutation.
import sys; sys.path.insert(0, "research/kit") from group_algebra import cyclic_product, build_2bga mul, _ = cyclic_product(62) HX, HZ = build_2bga(mul, [0, 59, 2], [0, 20, 12])
Equivalently a(x) = 1 + x^59 + x^2, b(x) = 1 + x^20 + x^12 in F_2[x]/(x^62 - 1), H_X = [A | B], H_Z = [B^T | A^T].