The default primitive for AI integration in 2026 is coupling. Push the model's shape into the customer's shape, or pull the customer's shape into the model's. Forward-deploy the engineers. Embed the agents. Normalize the data. Train on the domain. The framing is "integration"; the structural move is to make the customer's governance surface and the vendor's governance surface share state.
The cost of that move is invisible at the moment of integration. It becomes structural and expensive forever after. When the vendor changes their model, schema, pricing, or terms — and they will — every coupled customer carries the change. The customer's governance was never anchored to a stable thing.
Forge runs a different primitive. The rule we use is the Volatility Handling Law, and the rule has six layers. L0 is doctrine: what the federation governs, period. L1 is policy that applies doctrine. L2 is the practice operatives execute under policy. L3 is a contract snapshot, the observed current shape of an external thing. L4 is a probe, the instrumentation that reads external surfaces and converts what it reads into federation-native signals. L5 is the adapter, the thin connector between an external surface and a probe.
The promotion gate is explicit. An external system's specific schema, API field, model version, or endpoint is never doctrine. It's an L3 contract snapshot at best: a current observed shape, subject to drift. Only the handling law itself can be promoted to doctrine. The handling law specifies how the federation detects drift, verifies what crossed the boundary, contains drift inside the adapter, and recovers when the contract snapshot is invalidated.
The integration model that follows from this rule is: don't merge schemas. Deploy probes.
When the federation connects to an external system — vendor API, partner platform, customer data source, third-party model — we don't normalize their data into our ontology. We declare an observation boundary. The external system publishes whatever it publishes at known endpoints. An L4 probe reads that surface and converts what it reads into federation-native signals. Those signals enter the normal governance pipeline. Signal becomes warrant. Warrant becomes lesson. Lesson can refine the federation's ontology over time. The external system never enters the federation's doctrine.
When the vendor changes — and the relevant question is when, not whether — only the L5 adapter has to change. Every governance rule, doctrine layer, policy, and practice survives unchanged. The federation's shape was never coupled to the vendor's shape.
There is a related rule about vocabulary, and it's worth naming. Don't import vendor vocabulary into governance. Vocabulary creates monitoring obligations. If a law says "personality crystallizes at 48 hours," the governance surface watches the clock. If the same phenomenon is described as "attractor convergence at interaction-density threshold N," the surface watches the interaction counter. Both are law-form. One produces a working governance surface. The other breaks the first time an agent has dense interactions in hour three or sparse interactions through hour seventy-two. The vocabulary you adopt at integration time becomes the variable your governance surface monitors. Choose what to monitor; don't inherit the choice from a vendor whose ontology was built for different reasons.
I expect three pushbacks on this. They're worth stating directly.
The first is that this is just service-oriented architecture with more layers. SOA decouples message format from implementation. The Volatility Handling Law decouples governance from external shape. SOA still requires both sides to agree on a message contract; the volatility law treats even the message contract as L3, subject to drift, and provisions L4 probes that can survive contract changes. The two rules operate at different layers.
The second is that probes are slower and less rich than direct integration. They are. That's the point. The cost of coupling is invisible until the vendor changes; the cost of a probe is visible at integration time. A federation that wants to outlast any specific external system pays the visible cost on purpose.
The third is what if the external system never changes. Then the federation paid a small premium for nothing in that case. But the law has to govern the worst case, not the best one. Doctrine that assumes external shape will hold is doctrine that breaks the day it doesn't.
The frame: when you don't govern the shape, you don't govern the shape. You govern your own response to it. That's the only doctrine that survives the next vendor change.