← all posts
// analysis · runway

Runway Solaris renders working UI frame by frame: what breaks when there's no DOM

Runway announced Solaris on August 31 as the first "Interface World Model": a system that generates functional application and web interfaces frame by frame, in real time, in response to clicks and drags, with no HTML, CSS or JavaScript underneath. It is a research release in early access, and the weekly roundups for September 1-8 list it as the thing to watch. They are right to watch it, and most of them are drawing the wrong conclusion. Solaris does not mean software stops being code. It means there is now a second way to produce something that looks like software, and that second way breaks nearly every tool the industry uses to verify the first.

What is actually being claimed

A world model predicts the next state of an environment given an action. Point that at interfaces and you get a model that, shown a screen and a click, renders the next screen. There is no DOM tree, no layout engine, no event handlers; the "button" is a region of pixels the model has learned usually leads to a certain next frame. Runway's own study reports a 61% preference from 250 participants, and the fine print matters: the preference is for instruction-following, meaning the generated UI did what the prompt asked, not for production readiness, and it is a vendor study.

Runway itself lists what is unsolved: text legibility, coherence across long sessions, and accessibility. Those are not polish items. Legible text is the difference between a UI and a picture of one; long-session coherence is state; accessibility is the entire non-visual contract of an interface.

A user interface without a DOM is not an interface with less code; it is an interface with no contract.

What breaks when there is no DOM

Everything in the verification stack assumes structure. Playwright and Selenium assume selectors: a locator resolves to an element, the element has a role and a name, and a test asserts on it. With Solaris there is nothing to resolve. You are back to pixel matching, which the industry spent a decade escaping because it is brittle, slow and blind to meaning.

  • Functional tests. No selectors, no assertions on state, no way to know whether the Save region saved anything or just rendered a frame that looks saved.
  • Accessibility. Screen readers consume an accessibility tree. There is none. A world-model UI is inaccessible by construction until someone generates a parallel semantic layer, at which point you have reinvented the DOM.
  • Determinism. The same click can render a different next frame. Regression testing needs reproducibility; generative models offer distributions.
  • Audit and compliance. Regulated software has to show what the user saw and what the system did. A video stream shows the first and has no record of the second.
  • Security. There is no input validation layer because there is no input model; the backing state is whatever the model implied it to be.

The only test strategy that fits is the one we use for other generative output: judge the frames with another model against a rubric, the way output validation works for text. That gives you a probabilistic quality score, not a pass or fail on behaviour.

Where it makes sense anyway

Prototyping and exploration. A designer who can click through a generated app before any code exists gets faster feedback than any static mockup. User research, where the point is to observe reactions, not to ship. Concept validation for products that do not exist yet. Game-like or creative interfaces where visual surprise is the feature. In all of those, the absence of a contract is fine because nothing depends on the contract. The moment money, health, law or a screen reader depends on the interface, it needs to be code, or at least to be backed by code that Solaris paints over.

The honest limitation

Nobody outside early access has used Solaris, including me. The 61% figure is Runway's own study on its own criteria; there are no latency numbers, no session-length limits, no hardware requirements and no pricing in what has been published. A future version may well emit a semantic layer alongside the frames and make half of this article obsolete. Until it does, treat Solaris as a prototyping instrument with a striking demo, not as a signal that your Playwright suite is about to become irrelevant.

#runway#world-models#testing#ui