What was already known
Wu and Santhanam established eventual prediction for several matrix properties and asked the diagonalizability question. Their closed-set results alone do not settle it because diagonalizable matrices form neither an open nor a closed locus; the source cites a revised 2024 manuscript still listing it as open.
AI involvement
Zhao chose the problem and directed the work. He credits GPT-5.6 Sol Ultra through Codex with the proof strategy and counterexample, then revised the manuscript.
The model supplied the central proof strategy and counterexample, while Zhao led problem selection and revision; the disclosure does not credit the model with every proof step.
Importance
Answers the fixed-dimensional diagonalizability question with a broader semialgebraic prediction principle.
Limitations
The result is fixed-dimensional and does not claim an efficient predictor. A finite-r>1 moment extension and an integrable-case failure boundary are separate parts of the paper. No completed independent mathematical check is documented.
Technical verification
The v1 paper gives a self-contained proof, but no completed human or machine check of the model-origin strategy is documented; Reported is the evidence default.
Human understanding
Default for undocumented human understanding: the paper explains the predictor, but model-assisted exposition and human revision do not identify a human explanation or grasp of the central model-origin reasoning.
Correction and publication history
GPT 5.6 Sol Ultra supplied the strategy and counterexample; Zhao chose the problem and directed revisions.
Read this published revision