Four chapters. None of these is a market forecast: they're the assumptions I'm betting my time on, which means they cost me if they're wrong.
A model produces plausible output and can't prove why. As long as the cost of an error is badly written text, that's fine. The moment output moves money, denies credit or signs off a supply order, plausibility isn't enough: you need a chain of custody for the decision.
I don't ask a system to be infallible. I ask that when it's wrong, you can reconstruct exactly where.
It solved exactly one, but solved it well: getting parties who don't trust each other to agree, without an arbiter. Everything else — tokens, speculation, promises of universal disintermediation — was noise that obscured the one function that's genuinely hard to replicate elsewhere. That function now has an obvious customer: AI.
An agent operating in production records every decision: input, retrieved context, tools called, output, who approved it. That ledger gets anchored on-chain at intervals. The data doesn't go on the chain — the proof that the data wasn't rewritten afterwards does. It's a small difference to explain and an enormous one to implement.
Not in startups proving the concept on a clean dataset, but inside the .NET systems that have been running for twelve years and that nobody can switch off. That's where auditability isn't a nice-to-have but a survival requirement — and so that's where convergence stops being a thesis and becomes a quote.
Otherwise it would just be a well-written opinion.