Why it matters:

The Juice Is in the Books: Fed the same textbooks as in-context guidance, Transfer Par@8 reaches a full 100.0 on Qwen3-8B. The material provably suffices, and olympiad capability is recoverable from the corpus alone. What remains is not access to knowledge but the ability to extract it.

Absorption Without Transfer: The strongest self-evolution method roughly doubles Application Par@8, absorbing the hard problems inside the books. Yet on the harder olympiad problems it closes only about 7% of the gap to that guidance ceiling. Learning the coursework does not carry over to problems one step beyond it.

A Plateau, Not a Ladder: Every method we test tops out early and stays flat as more compute is committed, so more of the same loop buys almost nothing. The remaining gap is a method problem, not a data or compute problem.