1/ The so-called "AI settlement" is purely an overextension of engineering intuition. Traders are forced to price "second-order variables": event probability + model madness probability + attacker manipulation probability. Calibration directly pollutes, and liquidity evaporates instantly. 2/ Technological illusion: don't think that adding a greedy decoding model makes it deterministic. The non-idempotency of the inference stack means that: the same input under high load may produce different results. The settlement benchmark is floating, so what’s the point?