ros/claude-plugins-official
1
0
Fork 0
mirror of https://github.com/anthropics/claude-plugins-official.git synced 2026-03-20 11:33:08 +00:00
Code Issues Packages Projects Releases Wiki Activity
Files
c3f6d9e9fac9a5caa3967a1edbe20a905d138ad6
claude-plugins-official/plugins/math-olympiad/README.md
Ralph Furman 3c9bf4ff5d Add math-olympiad skill — adversarial verification for competition math 
The skill that addresses the Proof-or-Bluff gap: self-verified 85.7% IMO
becomes <5% under human grading. Uses fresh-context verifiers armed with
specific failure patterns (not generic 'check logic').

Validated: 17/18 IMO+Putnam 2025 solved, 0 false positives, 2 novel proofs.
See eval data in anthropic monorepo sandbox/sandbox/ralph/math_skills/.
2026-03-19 23:17:36 +00:00

1.2 KiB
Raw Blame History

math-olympiad

Competition math solver with adversarial verification.

The problem

Self-verification gets fooled. A verifier that sees the reasoning is biased toward agreement. arXiv:2503.21934 ("Proof or Bluff") showed 85.7% self-verified IMO success drops to <5% under human grading.

The approach

  • Context-isolated verification: verifier sees only the clean proof, never the reasoning trace
  • Pattern-armed adversarial checks: not "is this correct?" but "does this accidentally prove RH?" / "extract the general lemma, find a 2×2 counterexample"
  • Calibrated abstention: says "no confident solution" rather than bluff
  • Presentation pass: produces clean LaTeX/PDF after verification passes

Validation

17/18 IMO+Putnam 2025 problems solved, 0 false positives, 2 novel proofs found. See the skill's eval data in the anthropic monorepo.

Install

/plugin install math-olympiad@claude-plugins-official

Use

> Solve this IMO problem: [statement]

The skill auto-triggers on "IMO", "Putnam", "olympiad", "verify this proof", etc.