Take the frame loop at its word. Each frame depends on the frames already produced [5], and mouse and keyboard activity arrives as conditioning data for the next one [4]. That is the whole state model. Nothing in Runway's description names an addressable object, and the announcement ships no API [2], so the only record of what a user did is the pixel history. Regression-testing that means diffing images.
The training pipeline shows where it hurts. Solaris sits on Gen-4.5 and the world-model line Anastasis Germanidis introduced in December 2023 [3]. Runway distilled the video model's multi-step denoising into fewer steps, then trained the faster model on its own outputs specifically to improve stability during longer interactions [6]. Stability over time was expensive enough to earn its own stage, and Runway still lists extended-session coherence as open, along with stable legible text and grounding responses in verified information [16].
Now put that against the first demo. A shopper drags clothes onto an image of themselves [10]. That screen carries a price, and a price is text that has to read the same on frame 400 as on frame 1. Same for a stock count.
The "no code" framing does not survive the architecture. A separate language model interprets the request, decides whether to alter the current scene or move to another, and sends instructions to Solaris, while text prompts define what a click or drag means inside a scene [8]. Runway's own account leaves prompts, a language model and predefined starting material determining what happens [9]. The specification did not disappear; it moved into prose, where there is no type checker and no useful diff.
The preference study is the part to read slowly. Runway ran 250 participants over 30 examples for nearly 7,500 pairwise judgments, and reports 61% preference for Solaris on following instructions against 24% for a Claude Opus 5 coded interface built from the same starting image and request [12]. Natural behavior split 71% to 21%, with most of the remainder classified as equivalent [13]. Divide it out: 7,500 over 250 is about 30 judgments each, one per example [18]. The evidence unit is 30 interfaces. For that 61% to transfer, your screens would need to sit in the same class as those 30, your baseline would need to be one model's single-shot code from a screenshot, and your acceptance test would need to be a stranger's impression of whether the instruction was followed. Mine is whether the total is right.
The reconstruction test points the same direction. Across 30 interfaces rebuilt from screenshots by GPT-4o, Gemini 2.5 Pro and Claude Fable 5, measured with structural similarity and DINOv3 features, fidelity fell as the source images grew more visually complex [14]. That is a finding about screenshot-to-code, not a ceiling on Solaris.
Runway is straight about the scope, saying the results do not establish Solaris as a replacement for conventional software on reliability, cost or accessibility [15]. Good craft, honestly bounded. I would run one test before believing more than that: same starting image, same click sequence, twice, then diff the frames. Determinism is the guarantee a coded UI gives away for free, and this design has to earn it back.
