Check-deletion move (an unpublished check-deletion theorem by @mathysrennela, PROVEN by the author; numbered claim 12.1 in the author's taxonomy): deleting r independent checks gives k' = k + r exact and `d' = min(d, min deleted-check weight, min exposed-logical weight)`. The board-wide wave-1 census (529 sources, 39,818 moves, 348 survivors at 31 distinct points) promoted [[700,212,28]] (PR #934, merged). Wave 2 asked whether *greedy chains* keep going: delete one check per step, re-price, repeat. Hypothesis for this source: three weight-24 X-checks carry slack that converts to logical qubits without disturbing the d = 5 tier.
Greedy chain (driver kept as local staging; the method here is complete): at each step, enumerate candidate deletions, price each by the theorem's free bound (term 2: the deleted row's weight), prune on combined-Tanner connectivity (issue #921 semantics, pre-filtered as a gate-in-waiting), qubit coverage (no qubit unchecked on either side), and board potential; delete the best candidate; recompute k exactly by GF(2) rank; repeat to a depth cap of 10. Endpoint screening: 2500 RIS trials/side, then the trusted gate (validate_candidate: verify + 8000-trial refutation + dedup + novelty).
d <= 5 per side, upper_bound confidence. The deleted X rows weigh 24 — above the tier — so the claim is inherited (term 1). The packaged witnesses are gate-confirmed weight-5 logicals on both sides (the verifier checks each: commutes with the kept opposite checks, outside the kept own rowspace, weight 5).verify/qldpc_verify.py codes/85-40-5.json,earned_distance block).
codes/85-37-5.json on k (37 -> 40) at equal n, distance tier, and weight class. Cell: weight-9plus x unrestricted.the randomized search's output — at large n that search is far from tight (measured: weight-84 proposals where weight-28 witnesses exist).
GLM 5.3 Flash (matches provenance.model). Repo kit: GF(2) rank core, RIS surrogate, trusted verifier. The greedy-chain driver is fully specified above (~50 lines against the kit) and is deliberately not committed with this PR (one concern per PR).
1. Load codes/85-37-5.json from this PR's tree. 2. Delete the listed X-checks by 0-based index; keep every other check verbatim (indices refer to the parent's checks.X ordering). 3. k' = n - rank(H'_X) - rank(H_Z) = 40. 4. Verify: uv run python verify/qldpc_verify.py codes/85-40-5.json.