What caught me while working through OpenGradient — $OPG , #OPG , @OpenGradient — as a CreatorPad task was a quiet distinction the project seems to navigate carefully: verifiability of computation is not the same thing as verifiability of fairness. The architecture can cryptographically prove that a model ran exactly as specified — that the inference wasn't tampered with, that the output was genuinely produced by a given set of weights. That's meaningful. But what it can't prove is that the model itself, faithfully and perfectly executed, is treating different groups equitably. A biased model, run correctly, generates a verifiable bias. The proof confirms the integrity of the mechanism, not the justice of the outcome. I kept thinking about who benefits most from that distinction — probably developers and enterprise clients who can now say their AI ran "as intended," which is a defensible claim, just a narrower one than it sounds. The fairness claim gets displaced upstream, into the model's training, somewhere outside the chain. Whether that's a design limitation or a deliberate scoping of what a protocol can reasonably promise, I'm still not sure.