The claim entered through the common gate. The costly reader stood upstairs. No clerk could find the rule that said which questions earned the better chair.
Most requests go to the cheapest model likely to cope. A smaller share goes to the expensive model reserved for work where failure costs more. The routing rule looks like infrastructure. It is also an institutional theory of consequence.
Vercel's AI Gateway Production Index for July 2026 made the split visible in its June data. Open-weight models ran 29 percent of gateway tokens on under 4 percent of spend. Anthropic occupied the other end of the distribution, taking 61 percent of spend on 32 percent of tokens. Vercel described the pattern as high-volume work moving to low-cost models while high-risk work stayed on the frontier.
A classification disguised as plumbing
Every router needs a criterion. It may inspect length, task label, customer tier, latency target, or an estimate of difficulty. Somewhere in that logic, an engineer decides which requests deserve the costly model.
That decision is a materiality classification. It determines where the institution spends for greater capability and where it accepts a thinner margin. The router can make the choice thousands of times before anyone recognizes it as policy.
Rule 3 of the Interim Protocol already names judgments that deserve special care: eligibility, significance of impact, adequacy of mitigation, compensation, consent, and the scope of consultation. A cost router rarely receives that list. Its categories come from what the implementation can detect.
Succession moves the line
Replacing one model alters the relationship between cost and capability. The cheap model may become good enough for work previously sent upward. The premium model may lose an ability on which a high-impact workflow depended. Even if the routing code stays unchanged, the materiality line moves.
A model migration can therefore rewrite policy without a policy decision. In the constructed appeal, the October request reaches the same routing code that handled work in March. Its model choices have changed beneath it. If the router sends the case to a cheap tier because the input resembles an ordinary document review, the institution has reclassified the appeal before any reviewer sees it. The log may record only which endpoint answered.
Put the judgment in the register
The routing table needs an owner outside the infrastructure team. Its entries should state the consequence being protected, the evidence for the chosen model class, and the event that triggers review. Retirement and price changes belong on that list, as does a material change in measured behavior.
The record should preserve rejected routes as well as successful ones. When a consequential request goes to a cheaper system, reviewers need to know whether the route reflected an approved classification, a fallback, or a budget ceiling.
Somebody set the router. Succession planning begins by naming who, under whose authority, and with which account of what matters.
Companions
- The classification rule: Interim Protocol.
- The migration deadline: Pinned.
- The earlier account of asymmetry: The Scale That Only Runs One Way.
These notes come out of Sociable Systems, a practice that reads AI-shaped documents the way a hostile reviewer will, before a lender or a court finds the gap. The argument has an operational form: the Interim Protocol sets out four rules for AI use in environmental and social deliverables, covering disclosure at touch-point grain, evidence custody, the phrases no automated screening may settle, and a hostile read before anything ships. Free, and written to be cited or retired once institutional guidance arrives.