The article proposes a responsible AI‑math release framework after receiving over 600 community responses. It separates fully understood papers (2.A), which must be posted as preprints, submitted for peer review, and presented at seminars, from AI‑generated results not yet understood (2.B). For the latter, labs should (I) survey related literature and cite original sources, then prompt models to rewrite proofs in conventional, well‑structured style; (II) deposit results promptly in open repositories with persistent identifiers, allow comments, and provide metadata on model name, prompts, chain‑of‑thought, runtime and cost. Formalization is encouraged, using community‑standard artifacts (copyright header, challenge file, formalization.yaml) and machine‑readable metadata; any delay must be disclosed. Labs must document how AI was used for each result and, when multiple results are released, summarize the problem set and failures. To foster human understanding, AI labs should fund community‑run institutions that support workshops, summer schools, post‑doc positions or expository writing, with support matching result complexity. Finally, broad, equitable access to public models is required to prevent a two‑tier mathematical ecosystem.