Nebius × NVIDIA Global AI Hackathon · judge's guide

Question an AI proof, check it in Lean, and build on what holds.

MathCase keeps a proof attempt's story together: the claim, the AI's candidate, Lean's response, and the limits of what the saved evidence establishes.

Hackathon judging: Dec 9–23, 2026 · winners announced Jan 25, 2027 at noon Pacific time.

Try the Case Desk

Start with a real saved case.

Open either case to inspect its statement, step through the saved result, and try the exact-integer counterexample. Nothing dispatches a model or starts a proof run.

Open the saved Case Desk

Saved local calculation

A counterexample narrows the claim

A composite number can divide a square without dividing its root: the saved example uses 4 and 2.

4 ∣ 2², but 4 ∤ 2

It shows: this unrestricted statement has a counterexample. It does not show: a general theorem or a failed AI response.

Open this saved case

Public coverage gap

A third unrestricted case is not packaged yet

The current bundle has two suitable saved examples for this public walkthrough. Its other bundled case is not cleared for this view, and no replacement case with the same statement–failure–Lean evidence chain is available in the public dataset.

Next needed: a separately reviewed public case with its own statement and saved evidence. No substitute has been invented.

Not available in this public walkthrough

The saved reader preserves the original two public cases. Evidence file locations are shown in the reader; individual machine-readable records are not linked as the visitor experience.

How to read a result

A check applies to one exact statement.

For the prime-divisibility example, the saved record includes an actual Lean rejection and a repaired candidate. That makes the proof process inspectable without turning a known theorem into a discovery claim.

In plain language

If a natural prime p divides n squared, then p divides n.

p prime, p ∣ n² ⟹ p ∣ n

What it proves: the exact formal theorem was accepted by the saved Lean check.
What it does not prove: novelty, independent review of every explanation, or open-problem capability.

  1. 01

    The first candidate was rejected

    It supplied the divisibility assumption where Lean required the primality assumption. The failed attempt remains available in the saved case.

    Failure preserved
  2. 02

    The repaired candidate used both assumptions

    The saved proof applies the known result Nat.Prime.dvd_of_dvd_pow. Its recorded Lean check passed.

    Saved check passed
  3. 03

    The evidence stays in its lane

    A kernel check validates the formal proof for its formal target. It does not establish that the theorem is new or that a finite search proves a general conjecture.

    Limits stated

Read the result, failure, and saved proof in the Case Desk. The webpage itself makes no live proof claim.

See the wider evidence map

Latest research update · Nov 13, 2026

The Collatz work is growing beyond individual shards.

A chunked M20 composition compiled as a local candidate under the two-gigabyte cap. Its source and run receipts are recorded privately; a public reproduction package is still pending, so the site does not present it as a public replay result.

Read the status and limits

Beyond the judging walkthrough

Browse the work by question.