What the panel can actually do

Every seat — including the referee — gets the same tools. The referee has to be able to check a claim, not just weigh how it was argued.

Literature search

arXiv, OpenAlex, Crossref, Europe PMC and Semantic Scholar — the primary sources, not blog restatements of them. Maths, physics, CS, biology, medicine, economics.

Seats are told to check what is already proved before attempting it themselves, because re-deriving a known result wastes your money.

Web research

Self-hosted search with real site restriction — ask for arxiv.org and you get arxiv.org, not whatever the engine felt like returning.

Pages are fetched and read, so a citation can quote the passage it actually relies on.

Sandboxed code

Python with sympy, mpmath and numpy, so a seat can check a bound, search for a counterexample or verify an algebraic step instead of asserting it.

Runs with no network access, capped CPU and memory.

Machine-checked proof · live

For mathematics, a proof can be submitted to the Lean 4 kernel. The kernel either accepts the proof term or it does not — there is no room for a persuasive argument.

A `sorry` is rejected even though Lean only warns about it, and declaring your own axiom is rejected too. Every theorem is interrogated with `#print axioms` and must rest solely on the standard three.

Escalating instead of guessing

Every seat can also stop and ask for help. A stuck model normally produces another paragraph restating the last one — at full price. Here it names the single question blocking it and hands that to whichever seat is best placed to answer, along with what it already tried. The expert gets the question and the relevant established results, not the argument that produced the impasse, so being stuck costs less than arguing.

And the referee's own tools

Rule on a criterion, add one the goal implies, direct the next round, or end the match. A criterion cannot be marked met without evidence a third party could re-check — a citation with the passage quoted, a computation with its real output, or a kernel-accepted proof. That refusal is enforced in code, not asked for politely in a prompt.