The paper proposes freezing model weights and accumulating independently verified solutions in an addressable persistent memory. Its abstract reports that four architectures from four vendors achieved 180/180 on 180 fresh instances across nine problem families, with zero generation tokens and bit-exact deterministic reuse; an emptied-memory control solved nothing. It also claims 88/88 consistency-gated acceptances, 77/80 reasoning-method transfer, 1.4-microsecond memory selection, and a movable 6-million-token window on one 46 GB GPU. These unusually strong claims require full-paper inspection and independent replication.
No heat snapshots are available in the last 24 hours.