Runway's Solaris Removed the Code Layer. Ask What Your Auditor Diffs.

Runway ran a user study with 250 participants across 30 interaction examples, collecting nearly 7,500 pairwise judgments. Each judgment compared an interface generated by Solaris against the same interface produced as code by Claude Opus 5, both starting from the same image and the same interaction requests. For following the requested interaction, Solaris was preferred in 61% of comparisons compared to 24% for the coded result, with 13% rated equivalent. For natural behavior the margin widened: 71% against 21%, with 6% equivalent.

Both rows sum to 98%. Runway prints them that way and leaves the two missing points unexplained. I am reproducing them as printed rather than rounding them into a number the source never published.

For anyone accountable for software, the decisive detail sits in what Solaris declines to produce. “Every frame is synthesized as you interact, allowing the interface to respond continuously.” It leaves behind no HTML, component tree, stylesheet, or repository. The interface exists only while it is being watched.

Every Control You Own Assumes an Artifact

Walk your own release pipeline and name what each stage operates on. Code review reads a diff. Static analysis parses a syntax tree. Change history stores commits. Reproducible builds hash inputs and compare outputs. Accessibility testing queries a DOM. Every one of these takes an intermediate representation as input, sitting between what somebody intended and what a user finally sees.

Generation-to-pixels deletes that representation by design. Runway made that choice deliberately and is transparent about it. The consequence is that the governance apparatus fails outright rather than degrading. It has nothing to bind to. A reviewer has nothing to approve, since nothing was ever written down, and a regression test has nothing to assert against on a surface synthesized fresh at every interaction.

We argued in the case for world models as a governance layer that a model of the environment gives control systems something to reason about. Solaris runs the same class of technology in the opposite direction. The world model is the renderer, and the renderer keeps no minutes.

Accessibility Is the Consequence Already Written Into Law

Runway lists its own limitations: legibility of generated text, trust and hallucination in instructional contexts, coherence across long sessions, and accessibility and assistive-technology integration. Three of those are quality problems that will improve as models improve. The fourth is structural.

Screen readers ignore pixels. They read an accessibility tree derived from the DOM: roles, labels, focus order, live regions, relationships between a control and the text describing it. The accessibility regimes your buyers operate under, and the audits those regimes trigger, are expressed as assertions against that tree. WCAG conformance claims are made about markup. Automated scanners parse markup. Remediation tickets cite selectors.

An interface synthesized frame by frame has no stable tree to assert against. A button that exists as pixels in one frame was, in structural terms, a different button in the frame before it. Runway naming assistive-technology integration as an open limitation is the honest version of this. It is also, for a regulated buyer, the sentence that ends the procurement conversation.

Note what the published results cover. Runway reports preference on following the requested interaction and on natural behavior. Neither dimension speaks to whether someone using assistive technology could operate the interface at all. A 71% preference on natural behavior and an unresolved accessibility limitation sit side by side without contradiction. They are answers to different questions, and only one of them has a compliance officer attached.

Where This Actually Lands First

I doubt regulated production surfaces will move to frame-by-frame generation. I expect the pressure to arrive from a direction most governance programs are not watching: the prototype that stops being a prototype.

Runway states that interactions stop feeling interactive somewhere around half a second of delay, and the demo is responsive enough that people drive it rather than watch it. A designer generates a working interface, a stakeholder drives it, and it behaves better than the coded version in the majority of side-by-side comparisons. What happens next is the part I have never seen written down in a release standard. Somebody asks whether it can just ship.

The output governance argument applies with more force here than in any generated-code scenario, because generated code at least lands in a repository where existing controls can catch it. When the output is frames, there is no landing.

This is also the boundary between two different problems that get discussed as one. Agents authoring the design system is a review problem: the artifact exists, and the question is who reads it and against what standard. Frame-by-frame generation escapes that category entirely. It removes the object under review.

Do This Now

Write one sentence in your engineering standards that names the artifact requirement explicitly. Something close to: any interface reaching a production user must have a durable representation, versioned and inspectable, that renders it.

That sentence sits inert on the day you write it. Nearly every tool you use already complies. Its value is that it is already in place when someone brings you a working demo built without one and asks a reasonable question about shipping it. At that point the conversation is about an existing standard rather than an ad-hoc objection from whoever happens to care about audits.

Then run the second check. Ask whoever owns accessibility compliance in your organization what they diff when they certify a release. Get the actual answer: a DOM snapshot, a scanner report, a manual screen-reader pass, a signed statement referencing specific markup. Whatever they name is the artifact your pipeline cannot afford to lose. If nobody can name one, the interface question sits well down your queue.

Runway built something users prefer by a wide margin and published its own list of what it does not yet solve. That is more disclosure than most vendors offer. The work on our side is to know which of our controls the design assumes away, before a preference number that large makes the question feel rhetorical.

---

This analysis synthesizes Introducing Solaris (Runway, August 2026).

Victorino Group helps engineering organizations define which artifacts a governed release pipeline requires, before a generation tool removes one. Let’s talk.