Contingency-table sampling

The quantity is the ideal auxiliary-chain inverse-gap upper bound $A d^p$, where $d=10+(m+1)(n+1)\ge10$. Lower $p$ is better, with $A$ also tracked. This is an internal bound, not the complete algorithm runtime.

OpenAI baseline: $98{,}560d^{126}$, released Oct 6, 2:58 PM PT.

These results are Claimed: their authors published them, and we have not checked them.