Essay · The evidence layer

The Drift SLA

Your AI vendor contract covers uptime, latency, and support tickets, and says nothing about the model staying the same model. The missing clause, how to draft it, and the referee problem that makes it enforceable.

Open your AI vendor contract and read the service levels. Uptime, almost certainly: four nines, credits if missed. Latency, probably. Support response times, ticket severities, maybe a data-residency annex. Now look for the clause that governs the thing you actually bought: the behavior of the model. Whether it will act next quarter the way it acted in the demo. Whether it can be swapped, tuned, or quantized without you hearing about it. In nearly every contract I have seen described, that clause does not exist. You have bound the vendor to keep the lights on and left them free to change what the lights illuminate.

Why the clause is missing

Not conspiracy, chronology. SLAs were invented for infrastructure whose failure modes were binary: up or down, fast or slow. AI failure is neither; it is drift, a system that stays up, stays fast, and quietly becomes worse at your use case. Contracts lag categories. Vendors do not volunteer stability commitments because silent change is operationally convenient for them, and buyers do not demand what they have not yet learned to name. This essay is the naming.

Drafting the drift clause

A workable behavioral-stability clause has four parts. Identity: the specific model serving the contract is named and verifiable, not just branded, the problem I unpacked in The Model on the Invoice. Notice: material changes to the serving model require advance notice, with a defined window. Materiality: what counts as a material change is defined by measured behavior against an agreed baseline, not by the vendor's assessment of their own update, because the party making a change is the least qualified in the room to declare it immaterial. Remedy: notice violations and unacknowledged drift carry consequences, credits, exit rights, re-validation at vendor cost. None of this is radical. It is what every other supply chain already does about ingredients.

The referee problem

And here is where most drift clauses would die in negotiation: verification. The vendor will not accept your internal evals as the trigger, correctly, since your harness is your testimony. You cannot accept their telemetry, same reason in reverse. A clause neither side can neutrally verify is decoration. The clause becomes enforceable exactly when a third party measures the model continuously, against stable baselines, on a record neither side can edit, and both parties agree in advance to cite it. This is the quiet contractual role of an independent measurement bureau, and it is why I keep building Modelometer as neutral infrastructure rather than a buyer-side tool: a referee owned by one team is just another player.

Why honest vendors should want this

The counterintuitive part: the drift SLA is a gift to good vendors. If you serve what you sell and change it responsibly, a verified stability commitment is the cheapest differentiation available, because your quieter competitors cannot sign it. Markets where verification exists reward the honest disproportionately; that is the whole lesson of the assay office. Expect the sequence every standard follows: first vendors resist the clause, then one signs it and wins deals with it, then it appears in every RFP and resisting it becomes the red flag itself.

The negotiation this changes

Until your next renewal, you hold more leverage than you think. Ask one question in the room: will you commit, in the contract, that the model serving us will not materially change without notice, as measured by an independent record? Watch the answer carefully. Yes is a vendor you can build on. A carefully lawyered no is information worth every penny of the deal you did not sign. And if you are told no such measurement exists, that is no longer true, which is rather the point of this whole series.

One more clause worth writing while your lawyer has the pen: an audit-rights line for identity, stating that the buyer may verify, through an independent measurement service, that the serving model matches the contracted one, and that verification results are admissible between the parties. It costs a vendor nothing to grant if they are honest, which is exactly what makes the refusal informative. Contracts are instruments for surfacing information the other party would rather not volunteer, and in AI supply, identity and stability are the two facts most worth surfacing before you depend on them.