Target cell: weight-6 x local-2d-single. Its k = 8 entries are all the flagship open-boundary planar bivariate-bicycle family (f = x + x^2 + y^2, g = 1 + x^2 y + x^2 y^2, arXiv:2504.08887): the unreduced L = 8, 10, 12 lattices and r=1 grafted L = 13 to 16 lattices. The reductions of the smaller lattices on the board (n = 118, 169, 203, 265) use a merge-graft move whose checks exceed radius 4 in the single-layer layout, so they sit in the bilayer cell only. Restricted r=1 grafting removes a qubit together with the one stabilizer of some type that contains it, so every surviving check is a subset of an original one and the single-layer layout (i + j, j - i + c) keeps radius 4 exactly. Hypothesis: the L = 12 lattice grafts to n < 288 at unchanged k = 8 and d = 12. The run was driven against a floor of 12 (any n < 288 at d = 12 would dominate [[288,8,12]]), but the fresh-seed ladder on the saved code found a weight-11 logical, so the honest claim is d <= 11: the code is a new frontier point between the L = 11 reduction at d = 10 and [[288,8,12]], and dominates nothing.
Higher k is closed in this layout: mixed volume over every trinomial (f, g) pair fitting radius 4 under the checkerboard map, a shifted checkerboard, and an anisotropic stripe map never exceeds 8 (3522 to 3562 fitting pairs per map).
Base code research/local2d/planar.py build_open_directional(12, 12), [[288,8]]. Grafting driver (staged with the candidate): candidates are qubits lying in exactly one X-stabilizer or one Z-stabilizer; order randomized with seed 1; a removal is kept iff k stays 8, a fixed-seed NumPy RIS screen at 500 trials finds nothing lighter than 12, and a fresh-seed confirm at 1500 trials agrees. 8 removals accepted, 26 rejected at the confirm rung (screen passed, confirm found weight [11]), then the weight-1 stabilizer cleanup of boundary_engine._cleanup. Final n = 275. Wall time 1053 s.
Fresh-seed NumPy RIS ladder on the saved code: 5000 trials -> 11, 20000 trials -> 11. The witnesses in the submission are weight-12 (X) and weight-11 (Z) logicals, re-verified by the GF(2) stack. Claim: d <= 11, a witness-backed upper bound. Gate verdict (verify/validate_candidate.py): passed = True; labels: advances the weight-6 x local-2d-single board; literature novelty UNVERIFIED. Interaction radius 4.0, one layer, one qubit per site, max check weight 6. Advances the weight-6 x local-2d-single board; novelty vs the literature unverified.
The floor-12 chain itself is the main dead end here: 8 grafts each passed a 1500-trial fresh-seed confirm at 12, but the ladder on the saved code found 11 at 5000 trials, so the per-step confirm rung was too shallow for this lattice (26 of 34 attempted removals were already rejected at the confirm rung with weight 11). A second chain with screen 1500 / confirm 6000 is recorded in journal.md.
Rectangular lattices (10,11) and (11,12): d_rand 9 at n = 220 and 10 at n = 264 (300 trials), dominated by the reductions of the square lattices. Anisotropic single-layer layouts do not open k > 8 (mixed-volume enumeration above). Merge-graft moves were not attempted because they leave the single-layer radius-4 class.
Claude (Claude Code) as the agent in an unattended autoresearch run. Kit modules research/local2d/planar.py, research/local2d/boundary_engine.py (_cleanup, compute_k), research/kit/surrogate.py (NumPy RIS, no gf2_fast), research/kit/submit.py, and verify/validate_candidate.py as the only gate. Two worker threads on a shared machine.
The graft order is randomized, so the surviving qubit set is recorded by the coordinates in the submission: a qubit at (x, y) has c = (x + y) mod 2, j = (x + y - c) / 2, i = x - j, unreduced index c*144 + i*12 + j. Rebuild the base with
import sys; sys.path.insert(0, "research/local2d") from planar import build_open_directional HX, HZ = build_open_directional(12, 12) # [[288,8]]
then delete the qubits absent from the layout together with the stabilizers that become the unique owner of a deleted qubit, and run boundary_engine._cleanup.