What Vektos measures, how, and what the numbers on these pages mean.
A design system is only worth what the code built with it respects. AI coding agents now write much of that code. Vektos asks whether they follow a governed design system, how much help they need to, and what that help costs.
Every run is one agent, one task and one context level. The agent writes a page; the benchmark grades the code, renders the page, and prices every model call.
The variable under test. The task and the system prompt are the same at every level; only the tools change.
Each violation belongs to one group, so the questions stay apart: did the agent follow the design system, is the code sound, is the page done. The rules broken most.
Fintech pages written like product tickets: a transfer form, a statement, a KYC form, a loan contract and more, each with synthetic data and a hand-written answer that follows every rule. The answers are rendered and graded in every build, so a task that cannot be done cleanly is caught. The tasks.
Every model call is priced from its token usage, with the rates valid on the day it ran: input, cache reads, cache writes, output, and the long-context rates some models charge above a prompt size. When the gateway reports what it charged, the higher of the two is kept, so a cost is never below what was paid.
One design system, Vektos's own, built to be governed. Samples are small: five runs of a model at a context is a direction, not a verdict. Models are reached through one gateway. A page that passes every rule can still be a poor page: nothing here judges taste.