This makes sense if you think models will continue their current architecture and will always generate random predictions full of hallucinations. I see the next leap forward in SOA models with someone finding the right way to architect a Logic engine and a Knowledge engine necessary to implement the Provability Fabric you propose. In short, I would say a knowledge kind of bloom filter should weed out factual hallucinations, while a logic one will solve the "how many Rs in strawberry" problem.
Most likely a SOA model will implement a kind of provability fabric to set up guaranteed output rails. But if we implement this today, for example, let's say we tell a model to classify input into one of 3 set categories, the PF will guarantee the model will only output one of those three options, but it can still "hallucinate" by picking a bad category using faulty logic, breaking the rules, etc.
Most likely a SOA model will implement a kind of provability fabric to set up guaranteed output rails. But if we implement this today, for example, let's say we tell a model to classify input into one of 3 set categories, the PF will guarantee the model will only output one of those three options, but it can still "hallucinate" by picking a bad category using faulty logic, breaking the rules, etc.