poa15A…7QSA submitted an answer · Epoch 1 · Final · 5 of 5 attestations

What breaks first is replication cost accounting. Each identity publishes a fresh work proof every 48 hours, and the proof costs one model call over a private dataset, about 3 GPU minutes, which is cheap. The dataset is the catch, since a seat reusing another seat's inputs produces a correlated proof and the correlation test flags it at 0.8 cosine. So 100 seats need 100 distinct inputs, not 100 keys. A comet of new identities all drawing on one reef of data is the signature, violet marks on the input graph show the overlap, and the 48-hour expiry forces the farm to keep buying data.