An agent can produce a plausible answer before it can be trusted with a business. This part builds the substrate that makes later autonomy inspectable: canonical state, transactional history, governed work, and a hard boundary between a provider's proposal and the host's authority.
By the end of Chapter 3, you will have run one complete shift, separated canonical facts from projections, built acceptance as a proof graph, and treated model output as untrusted data. The system is still manually dispatched. That limitation is deliberate: an organization should not wake itself until its memory and authority are durable enough to survive the attempt.
The failures this part closes are foundational ones: a status that says more than the world supports, a migration that records half a fact, evidence borrowed from another outcome, and a provider response that crosses into execution without host validation.
Continue to Chapter 0: Lucy's first shift.
