# Expert review checklist

For each candidate of interest:

1. Load `Hx.mtx.gz` and `Hz.mtx.gz`; independently verify GF(2) ranks, `H_X H_Z^T=0`, and `k=n-rank(H_X)-rank(H_Z)`.
2. Recompute row/column weights and Tanner connectivity if relevant to the hardware model.
3. Replay each included low-weight logical witness: check commutation with the opposite checks and non-membership in the same-type stabilizer row space.
4. For a `FORMAL_OPEN_LOWER_BOUND` item, independently rerun the low-weight exclusion and then extend it toward `target.required_distance`. Do not infer distance from BP/OSD.
5. For exact audits, independently solve every logical sector/direction or use a second exact implementation. The included witness alone proves only an upper bound.
6. Replay novelty against an independently curated literature registry. A `novel` label here is relative to the pinned repository registry.
7. Report exact integer evidence: `k*d*d`, `12*n`, solver/version, timeouts, matrix hashes, and the candidate ID.

Priority review question: does any formally open, registry-novel candidate reach its required distance under an independent exact solver? The current dataset makes no affirmative claim.