The AI Readiness model

Not quality. A system can have beautiful components, a well-argued token scale and a Figma library everyone loves, and still score badly here. Readiness measures something narrower: whether a competent agent, given only what your system publishes, would pick the right component, use it correctly, and know when not to use it at all.

That distinction matters because the gaps are usually invisible to the people who own the system. You know that Modal is for destructive confirmations and Drawer is for filters. You know it because you were in the room. None of that is written down, so none of it is retrievable, so an agent picks by name similarity and gets it wrong about a third of the time.

Take Semantics, the heaviest at 20%. For every component Onefold has discovered, it asks whether the system explains the component's purpose in a way that distinguishes it from its neighbours. A description that restates the name adds nothing:

“Modal component.” scores zero. “Use Modal for short interruptive tasks that need an explicit decision before anything else continues.” scores full marks, because it tells an agent both what the component is and, implicitly, what it is not.

The component score is the evidence-weighted average across components, and the workspace score is the weighted average across dimensions. Every level of that is inspectable. If Semantics is 44%, you can list exactly which components dragged it down.

The score is recalculated, never stored

A readiness assessment is derived from persisted analysis data, so the same evidence always produces the same score, and improving the evidence changes the score predictably. Write a usage guideline for six components and re-run: Semantics moves, and you can see it move. That is the feedback loop the product exists for.

• Whether your design system is good. Visual quality, taste, and whether the component set fits your product are not measured and should not be.
• Whether your team follows it. Adoption is a different problem. Token compliance touches the edge of it, but readiness is about the system as published, not as practised.
• Whether agents actually succeed. Not yet. Readiness measures the conditions for success. Measuring the outcome means running agents against real tasks and scoring the result, which is on the roadmap and will change these weights when it lands.

Because of what fails first. In practice, an agent that cannot tell two similar components apart produces the most damage: it invents a third one, and now the system has drifted. Accessibility expectations matter enormously, but they are more often absent from the implementation than from the documentation, which makes them a smaller share of what readiness can observe.

These weights are a starting position, not a law. Dimensions are stored independently of the weighting, so the model can be re-weighted without re-analysing anything, and any change will be published rather than applied quietly.