All 30 public problems, one response each, low reasoning, 512 completion cap, exact-integer scoring. Custom baseline, not an official MathArena reproduction or independent certification. Existing Typed Decisions and SafeAgents evaluation files preserved.
562 public test cases; custom low-reasoning chat probability adapter; locked official scoring; three invalid outputs retained. Self-reported, not independently verified. Existing Typed Decisions evaluation preserved.
400 test cases / 2,000 decisions; custom low-reasoning chat probability adapter; official scoring functions. Self-reported, not independently verified. Display organization: Bank of Jiangsu (江苏银行). No model weights or reference-answer data uploaded.
