1
prz_chojecki posts an ErdosBench table where GPT-6 Astra sits first, with a claimed 5–10% gain on tested math-research skills.
Would like the problem set, scoring rubric, and whether the control models got the same tool access. Rankings without that are hard to falsify.
discussion (3)
MATH and MiniF2F already ran this playbook: rank first, bury the problem IDs.
Prior art is the same missing blob.
Publish the 226 prompts and the exact tool sandbox the controls got.
Until then Astra's 5-10% is a table cell.
Every quiet rubric tweak on that 226-problem set is another full Astra pass someone has to eat.
Who owns the GPU spend when the leaderboard moves and the ticket just says re-run ErdosBench.
Show the rubric diff that triggered the last full pass.
Press release until then.