Typesafe's Jev is the fish at the poker table

The thread raises exactly that objection and never settles it. A solved cash game has one correct action per spot, while a support ticket or code review often has no ground truth. The linked article holds the data, not the discussion.

Barely. The captured thread is a one line compliment, so no methodology, counter arguments, or replications appear here. Any criticism of the eval's setup has to come from the article itself.

No, and the thread does not try. That figure is top-action agreement on a single solved board, not expected value loss, and the article's own caveats say run to run probabilities move. It is a smoke test, not a verdict.

Yes, adapted to your own domain. The transferable part is the method: solve the ground truth, precompute anything arithmetic, then check whether failures cluster on trap shaped spots. The thread documents no tooling.

It is backnotprop's TexasSolver evaluation of Jev's action choices, trap spots and decomposed question shapes. The thread's story URL points to backnotprop.com/blog/jev-poker/.

backnotprop published it on backnotprop.com/blog/jev-poker/, surfaced through the Hacker News thread titled Typesafe's Jev is the fish at the poker table.

The article's own caveats say its run-to-run probabilities move, so a single 63% top-action agreement on one solved board is not a stable score, and comparing runs needs the same boards and settings.