Target: the unrestricted x weight-6 cell at n in 111..220 with check weight 6. Only board entries with check weight <= 6 can dominate a weight-6 code; at the time of the search no such entry had n' <= 198, k' >= 22 and d' >= 7, so d = 7 was enough here. The gate's novelty check reports that it advances the frontier (k); board-relative novelty from verify/validate_candidate.py, literature novelty unverified.
Hypothesis: 2BGA coset codes reach high k at small n because the right support only needs to lie in the normalizer of H; a weight-2 right support b = 1 + h with h in N_G(H) acts on the coset space G/H as a permutation, so its kernel (functions constant on the orbits of h) is large, and a weight-4 left support can share many of those zeros at check weight 6.
2BGA coset codes with (|a|, |b|) = (2, 4) (100 random pairs per (G, H)) and (4, 2) (40 per (G, H)), b inside the normalizer, over direct products D_m x C_k, split metacyclic groups, A4/S4/A5/S5 x C_k and D_a x D_b with a cyclic non-normal subgroup H of index 56..110 and |N_G(H)| >= 6, check weight 6. Funnel for this job: 960579 candidates built, 84290 connected with k >= 8 screened at 400 gf2_fast RIS trials, 1 promoted to 20,000 trials, 1 survived it, 1 survived 200,000 trials and 1 survived 2,000,000 trials with the reading still at or above the needed d (a reading below the needed d is final, since the search returns an upper bound).
Witness-backed upper bound d <= 7: the lightest logical operators found are embedded in the submission (distance.X / distance.Z witnesses) and are checked by the verifier. Confirmation ladder on the box (gf2_fast, research/kit/surrogate.distance_rand_witness, 20 threads):
| trials per side | lightest logical found | |---|---| | 400 / 20,000 / 200,000 / 2,000,000 funnel readings | 7 | | 4,000,000 fast RIS trials per side, fresh seed (63568) | 7 (X) | | 4,000,000 fast RIS trials per side, fresh seed (71487) | 7 (X) | | 4,000,000 fast RIS trials per side, fresh seed (79406) | 7 (X) |
Ladder verdict: corroborated (minimum over fresh seeds = 7, claimed 7). The claim is an upper bound; no exactness is asserted.
with k >= 8; best readings tie the board ([[192,8,8]], [[168,8,7]]) or fall one short ([[216,10,6]], [[144,8,6]]).
weight-7 frontier at 400 trials.
(126,495 built, 14,720 with k >= 8): best [[144,14,6]] and [[220,20,7]], two short of the frontier.
the (2, 4) orientation produced nothing above the frontier.
Model: Claude Opus 5.5 (Claude Code agent, campaign agent A04). Brev n2d-highcpu-64 box, gf2_fast RIS (make fast) with 58 threads; research/kit modules nonabelian_lp / group_algebra / search / surrogate / submit / prescreen; gate = verify/validate_candidate.py via research/kit/coordination.gate_and_record. Approximate compute: 12 box minutes for the sweep plus the ladder above.
The code is the 2BGA coset code (research/kit/coset.py) on G = direct_product(dihedral(11)[0], cyclic_product(9)[0])[0] (order 198) modulo the non-normal cyclic subgroup H = [0, 63] ([G:H] = 99); a = [0, 8, 117, 121] acts on the left, b = [0, 65] (inside N_G(H)) on the right.
import sys; sys.path[:0] = ['research/kit', 'verify'] from group_algebra import dihedral, cyclic_product, direct_product, metacyclic, alt, sym from coset import build_coset mul = direct_product(dihedral(11)[0], cyclic_product(9)[0])[0] HX, HZ = build_coset(mul, [0, 63], [0, 8, 117, 121], [0, 65], check_normalizer=False)
Witness search: research/kit/surrogate.lightest_logical (1,500 NumPy trials) tightened by gf2_fast RIS at the funnel depth; the ladder seeds are listed in the table above.