Target cell: weight-8 × unrestricted. Following the construction family of arXiv:2609.06572 (BB-type codes with weight-4 generator polynomials, i.e. weight-8 checks), the sweep targeted the frontier gaps that paper left open: grids it never searched ((18,4), (24,3), (36,2) at n=144), asymmetric generator weights (3+5, 2+6 — still weight-8 checks), and specific (k,d) gaps computed against the board's weight-8-cell Pareto frontier, in particular [[144,8,>=13]] between [[144,6,15]] and [[144,10,12]].
13,200 constant-term-normalized random pairs (the normalization is complete per arXiv:2609.06572 Cor. 3.9) across 13 grid/weight configurations, ~59.5k samples after degenerate rejection, screened at 300 RIS trials (400 records kept), then 169 shortlisted records deep-screened at 20,000 RIS trials. The sweep was a single throwaway script (random pair sampler feeding the kit's screen); its full record lists are local working output and are not committed — the method is fully described here and the winner is rebuilt from the polynomials below. Seeds: numpy default_rng(20260909), screen seeds 11 and 23.
Confirmation ladder for this code (per side unless noted):
lightest logical found is weight 15 on both X and Z, nothing lighter.
fresh-seed RIS refutation found no lighter logical; dedup found no exact or WL-equivalent board entry; computed board-advancing in the weight-8 x unrestricted cell.
Claim precisely: witness-backed upper bound d <= 15. Two independent search mechanisms agree at 15 with nothing lighter; exact certification (MILP) was not run, so the d= tier is not claimed.
deep-screen survivors within this budget; the n=72 targets ([[72,10,>=11]], [[72,12,>=9]]) stay open.
the gap targets except the four [[144,8,15]] finds; dominated candidates were not staged.
Model: Omen Alpha 1.0 (opencode agent), search harness written and run by the model. Repo tooling: research/kit/bb.py (build_bb), research/kit/css.py (k recomputation), research/kit/search.py (screen, gf2_fast backend), research/kit/distance.py (BP+OSD decoder distance), ./qldpc submit (witness extraction, verification), and the trusted gate verify/validate_candidate.py. Local machine, single node, ~40 min of screening.
Rebuild (H_X, H_Z) with the kit on the torus Z_18 x Z_4 (n = 2*18*4 = 144):
from bb import build_bb
HX, HZ = build_bb(18, 4,
A_terms=[(0,0),(1,2),(9,3),(14,1)],
B_terms=[(8,2),(11,1),(12,0),(12,1)])
Then ./qldpc submit <npz> --authors @MathysRennela --family bivariate-bicycle. The construction string in the JSON carries the same polynomials.