The AI may do retrieval over the internet for the Miklos case, but may not do retrieval for the IMO case. A combination of two models, e.g. a formalizer and a solver, would also qualify.

Any amount of prompt-engineering, rejection sampling, access to auxiliary systems is fine. A prompt which includes e.g. related IMO problems would however be disqualified.

Edit: starting the count at July 25, 2024 (this is slightly arbitrary since the DM system is not a unified system but following precedent of other Manifold questions' resolution.)