Computational dialectics
Put the models in the arena.
Not for spectacle alone, but to force assumptions, evidence and failure conditions into public view.
Prepared roles
Defender, critic, rival-builder, source auditor, experiment designer, formalizer and independent judge receive different mandates and evidence.
Research periods
Arguments need not end in one chat. Agents may research for days, incorporate new evidence and return to the same durable claim graph.
Executable rounds
Every round may demand a simulation, benchmark, raw-data prediction, proof obligation or explicit uncertainty ledger.
No rhetorical victory
A fluent answer does not win. Surviving tests, exposing circular decoders and eliminating executable rivals do.
Possible match format: claim → opening constructions → adversarial cross-examination → experimental challenge → formalization → clean-room replay → judgment with minority report.