dsh-math-olympiad

When LLMs tackle olympiad-level math, the typical failure is not incompetence but overconfidence: a wrong answer delivered with no self-check and no admission of doubt. dsh-math-olympiad turns the solve-and-verify loop into a strict multi-stage pipeline: pure-reasoning solve, think-chain stripping, a fresh-context subagent that attacks the proof clause by clause, up to two repair rounds, and an honest no confident solution output with a blocker note when the proof still fails. Structural guarantees keep the verifier from ever seeing the raw chain of thought, eliminating self-review hallucination.

At runtime the skill follows Pólya-style reasoning, the verifier runs predefined failure-pattern attacks including open-problem specializations, and confidence is restricted to high, medium, or a candid abstention. When a LaTeX engine is available the proof compiles to PDF; otherwise it falls back to Markdown math without breaking.

Built for IMO, Putnam, USAMO, and AIME contestants and coaches, as well as researchers who need a solvable answer plus an auditable verification trail plus a clear confidence label. It does not promise a correctness rate, but it does promise rigorous process and honest output.