More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    Proof Checker: Mathematical Verification

    1

    Adversarial proof verification: an independent reviewer hunts the gaps, you fix them with full derivations, it re-checks until the argument survives — audit trail included.

    $5

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned
    Proof Checker: Mathematical Verification

    Proof Checker: Mathematical Verification

    Example session with this skill installed

    Check this proof before it goes in my paper. Claim: every bounded increasing sequence of rational numbers converges to a rational number. Proof: let (a_n) be increasing with a_n ≤ M. Let L be the least upper bound of the terms. For any ε > 0, some term lies within ε of L, and by monotonicity all later terms do too, so a_n → L. Finally, L is rational because it is the limit of rational numbers.

    • Read your context and instructions
    • Compiled the proof checker

    Cold review (independent reviewer, round 1) returned two findings:

    [FATAL] Final sentence — Invalid. "L is rational because it is the limit of rational numbers" is asserted, not proved — and it is false. Counterexample from the counterexample pass: a_n = the n-digit decimal truncation of √2 (1, 1.4, 1.41, 1.414, …). The sequence is increasing, rational-valued, bounded by 1.5, and converges to √2, which is not rational. The claim as stated falls.

    [CRITICAL] "Let L be the least upper bound" — Hidden assumption. Over ℚ, a bounded set of rationals need not have a rational least upper bound; the truncation sequence above has none in ℚ. The proof silently uses completeness, which ℚ does not have. Same gap as the fatal finding seen from the other side: this is a correct argument about ℝ wearing a ℚ statement.

    The rest checks out: the ε-step is valid wherever a least upper bound exists, quantifier order is right, no interchanges are used.

    Fix round: restate the claim over ℝ. There the least upper bound exists by completeness, the ε argument goes through unchanged, and "converges to a real L" is exactly what the proof establishes. The reviewer re-checked the revised statement; both findings closed, and a second counterexample attempt (decreasing sequences, two-sided bounds) found no new gaps.

    VERDICT: FAIL as stated (rational limit); PASS for the real-valued restatement — which is what the argument actually proved all along.
    WHAT THIS AUDIT DOES NOT ESTABLISH: rates of convergence, or anything about the paper's other claims.

    Connects securely to your tools. The creator never sees your data.

    About this skill

    A proof usually fails where its author stopped looking: the "clearly" that hides a lemma, the limit swap nobody justified, the completeness assumption the theorem statement never made. This skill runs an adversarial audit protocol built around one invariant — never grade your own work. An independent reviewer (a different model family, or a human expert) reads the frozen proof cold and returns findings graded by status (Invalid, Unjustified, Understated, Overstated, Unclear) and impact (fatal down to minor). You repair each finding with a full derivation, never a patch sentence, and the same reviewer re-checks the fix — threaded rounds, capped at three, with a decision table for when the trend says "restructure, don't patch." The reviewer also runs a counterexample pass on key lemmas, trying to break them before a journal referee does. The skill includes the eight gap patterns that account for most findings (handwave, illegal interchange, quantifier slip, hidden assumption, one-direction equivalence, incomplete case split, untracked parameter, misapplied borrowed theorem), an author self-check checklist and readiness rubric to run before you spend a review round, a complete audit-report template, edge-case handling for long proofs and reviewer disagreements, and a pre-delivery quality gate. The worked example shows a critical interchange gap and a dimension-dependence gap closed in round two — with the claim honestly weakened where the proof demanded it.

    • A PASS means the argument as stated survived adversarial review; it is not a formal, machine-checked verification.
    • The audit covers the mathematical argument only — empirical claims in the same paper need a separate review.
    • Review quality depends on the independence and strength of the reviewer you supply; a self-review is explicitly labeled weak evidence.
    • Very long proofs are audited lemma-by-lemma along the dependency graph, so full coverage takes multiple rounds.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before any payment.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Money-back guarantee

    Doesn't do what its page promises? You get your money back.

    Trust & safety

    Security scan

    Security scan passed 24 days ago

    Permissions

    No special permissions declared or detected

    Listed24 days ago
    Updated21 days ago

    What's inside

    Frequently Asked Questions

    That voids the method. The checker's value comes from independence — a different model family (or a human) that doesn't share the drafter's blind spots. If only the drafting model is available, the protocol has you label the result a self-review and treat a PASS as weak evidence.

    Graded findings, not vibes: each gap gets a status (Invalid, Unjustified, Understated, Overstated, Unclear), an impact (fatal to minor), a location, and — after your fix — a confirmation that the gap closed. Plus a counterexample log and a verdict against a fixed acceptance bar.

    No — that's the audit working. A fatal finding with a concrete counterexample (like a rational sequence converging to √2) saves you from a referee finding it. Usually the fix is an honest restatement: stronger hypotheses or a weaker conclusion the proof actually establishes.

    New in Research & Analysis