Models commoditize. Each generation is cheaper, more capable and more interchangeable than the last, and the enterprise that bet its advantage on a particular model has already watched that advantage depreciate. What does not depreciate is the organization's knowledge of itself: its architecture, its decisions, the evidence behind them, and the capacity to put that knowledge in front of a system at the moment it matters. That knowledge compounds. Every decision recorded, every dependency mapped and every control evidenced makes the next question cheaper to answer and the next agent safer to deploy.
This is the argument for treating governance as a layer rather than a function. A function is staffed; a layer is built. A function produces documents; a layer produces answers. A function is consulted when someone remembers to; a layer is consulted by every system above it, every time, because the systems are built on it.
#Why the adjacent categories cannot own it. Four categories sit near the control gap, and each has a reason it cannot close it.
Governance, risk and compliance platforms hold obligations, controls and attestations. They are the system of record for what must be true and whether someone has signed to say it is. They do not hold the architecture, and they do not compute from it. Their posture is declared, not derived.
Model and machine-learning operations platforms govern the lifecycle of a model: its data, its training, its evaluation, its deployment. They stop at the model's boundary. The enterprise the model acts within, with its systems, its decisions and its obligations, is outside their frame.
Agent frameworks orchestrate actions: they route tasks, call tools and manage memory. They are built on the assumption that the governance question has been answered somewhere else, and they offer a place to plug that answer in. They do not produce it.
Model providers govern the behavior of their own models, which is necessary and is theirs to do. They do not know your organization, cannot see your architecture, and should not be the authority on your intent.
Each of these categories is necessary. None of them holds the layer where intent, structure and evidence meet, and none of them is positioned to: a provider cannot be the governor of its own customers, an orchestrator cannot be the judge of its own actions, and a system of record cannot compute what it was only built to record.
#What the durable layer has to be. If governance is the layer that outlasts any model and any vendor, four requirements follow. It has to be computed from evidence rather than surveyed from opinion, or it is stale on arrival. It has to be linked to that evidence end to end, or it cannot be audited. It has to be consultable by agents as readily as by people, or the systems that most need it will be the ones that cannot reach it. And it has to be certifiable in a form a stranger can verify, or every claim it makes is a claim the buyer has to take on faith.
The category that meets those four requirements did not exist as a category until recently. The next chapter defines it.