A Surface That Moves Cannot Be Glanced At

The Slowest Interface in the Market recorded that three memos had established the Steward's interaction surface as load-bearing and that none of them said what form it must take, or why one form and not another. The chapter that followed answered a narrower question. The Steward Speaks. The Steward Sees. The System Reads. assigned each route between a Steward and an agentic system to the receiver's fastest channel, and put verification on vision because vision is the fastest thing a human has. The chapter after that placed the agent's boundary in the architecture. The form question is still open, and it has become urgent for a reason the corpus did not anticipate: the industry has started answering it, and the answer it is converging on is that the interface should be generated by the model each time it is needed.

This memo opens the form chapter by naming what that answer does to the one route where form is the entire function. Layout Drift is the condition in which the Steward's surface changes the position, form or meaning of its elements between reads — by generation, adaptation or personalisation — so that each verification becomes a search rather than a glance, and the verification route is downgraded from the Steward's fastest input to their slowest.

What the industry built

On 18 November 2025 Google Research published its implementation of generative UI, in which the model "generates not only content but an entire user experience", with interfaces "automatically designed and fully customized in response to any question, instruction, or prompt". The post describes the work as "a first step toward fully AI-generated user experiences, where users automatically get dynamic interfaces tailored to their needs, rather than having to select from an existing catalog of applications". The same day the capability shipped in the Gemini app as two experiments and in Search's AI Mode. Five months earlier a Google Developers post had described a research prototype in which every screen of a simulated operating system is generated by the model on each interaction, and stated the property plainly: "visiting the same folder twice could produce entirely different contents". In December Google made its A2UI format public. A2UI is not the same artefact: it is a declarative catalog, in which the agent requests "generative or template-based" interfaces built from "a set of common components" the host application renders, so that agents can "generate the interface which best suits the current conversation". A Gemini Enterprise product manager is quoted describing agents generating "the dynamic, custom UIs needed for any task, from data entry forms to approval dashboards". The two exhibits sit at different points on the same line. Google Research generates the canvas. A2UI fixes the components and leaves the agent to choose which ones appear, in what arrangement, per conversation.

None of this is a mistake on its own terms. The research paper reports that when generation speed is ignored, raters preferred generated interfaces to standard model output, and the use cases in the announcements — explaining a topic to a child, planning a trip, laying out a gallery — are consumption. A reader who asks one question and receives one interface loses nothing when the next interface is different, because there is no second read to compare against. The memo's claim is not that generated interfaces fail at what they were built for. It is that the property that makes them work for a single read is the property that breaks the one surface built to be read again and again by the same person at speed.

Why the verification route needs a still surface

The Steward Speaks. The Steward Sees. The System Reads. put verification on vision because a human takes in a rendered state faster than they take in the same state read as text or heard as speech. That ranking has a condition attached that the memo did not need to state and this one does. Vision is the fastest input only when the viewer knows where to look. A glance is a movement of the eyes to a known place and a comparison of what is there against what was expected. A search is the same movement repeated until something is found. The first takes a fraction of a second; the second takes as long as the layout requires, and the viewer does not know in advance how long that is.

A surface that rearranges itself converts every glance into a search. The cost is paid per check and multiplied by the number of checks in a session, and it surfaces as a rate: the fraction of the states the Steward is expected to verify that they verify in the time they have. A catalog of fixed components does not remove the cost, because the components are fixed and their composition is not; a form that is always built from the same three parts, in a different order each time, is still a form the Steward has to read. The Steward returning to check whether the line commitment they confirmed an hour ago is still standing does not know where the commitment is now rendered, whether it is rendered in the same form, or whether the element that showed it before means the same thing now. They read the surface instead of glancing at it. Reading is the human's slowest input, and The Steward Was Not Slow. The Steward Was Blind. established what happens when the verification route delivers state through the slowest channel: Verification Blindness, the condition in which a Steward on the verification route cannot perceive the system's current state at the tempo the system changes it. That memo located the cause in modality — state delivered as text or audio when it should be rendered. Layout Drift produces the same condition through a rendered surface, because a rendering the Steward has to search is a text they have to read, with pictures.

The cost then follows the path the corpus has already drawn. A Steward who searches the surface every time verifies less of it, verifies it less frequently, and stops verifying the parts that take longest to find. The Steward's Blind Spot named the reading that produces: Nominal MTTI, a long time between interventions because the Steward has stopped engaging with the surface, indistinguishable in the metric from a system that has earned its autonomy. The Arbitrage Has a Twin added the tempo. A surface that costs more per verification widens the Governance Tempo Gap from the Steward's side, at exactly the point in the architecture where the corpus has argued the gap can only be closed by changing what the Steward has to perceive. A generated surface changes what the Steward has to perceive on every read, in the wrong direction.

The control room learned this first

Process industries reached this conclusion decades before agents existed, under conditions close to the Steward's: one operator, a system that changes state faster than a person can read a log of it, and a cost of a missed abnormal condition measured in physical damage. The purpose clause of ANSI/ISA-101.01, the standard for human-machine interfaces in process automation, states that following it should "improve the user's abilities to detect, diagnose, and properly respond to abnormal situations". The whole apparatus of the standard — a written interface philosophy, a style guide adopted before the first display is built, a fixed hierarchy of display levels, colour reserved for directing attention — exists to keep the surface still so that the operator can see what moved. The standard was written for a screen a human sits in front of for a shift. It was written against exactly the instinct the generative approach now industrialises: that a richer, more tailored display serves the operator better than a stable one.

Arco takes the control room as precedent, not as template. The Steward's surface is not a distributed control system and the Stewardship Model is not a plant. What carries over is the finding, reached by an industry with a strong incentive to get it right: a person verifying a fast system needs the system's state to have a fixed home.

What may change and what may not

The distinction the arc will hold throughout is between the value and the place. The state a surface shows changes constantly; that is the point of showing it. What must not change is where each state lives, what form it takes, and what each signal means. A commitment count updates on every allocation. The element that displays it does not move, does not change shape, and does not switch from meaning "commitments standing" to meaning anything else. Adaptation of content is the surface doing its job. Adaptation of layout is Layout Drift, and it is drift in the corpus's usual sense: a change the Steward did not commit and cannot see, because the surface that would show it is the thing that changed.

The generated approach is a legitimate answer to a question the Steward's surface does not ask. It answers what interface best serves this prompt. The catalog approach answers a better question — which fixed components serve it — and still leaves the two dimensions that matter for a glance, arrangement and position, to be decided per conversation. The Steward's surface has to answer what interface best serves the ten-thousandth glance from the same person under time pressure, and the answer to that is the one they glanced at the nine-thousand-nine-hundred-and-ninety-ninth time.

The Operator's Verdict

An operator building a Steward surface in the next two years will be offered tooling that generates it, adapts it, or personalises it per session, and the offer will come with evidence that people prefer the result. The evidence is real and it measures the wrong route. Preference in a single-read setting says nothing about the verification route, where the surface is read thousands of times by one person whose picture of the system has to keep pace with the system.

The Stewardship Model's autonomy metrics depend on a Steward who is actively verifying. Anything that raises the cost of a verification lowers the rate, and a surface the Steward has to search raises that cost on every read. Whatever else the form chapter specifies, it starts from this: the verification route needs a surface that stays still, and the model does not get to redesign it.

Technology changes what is possible. What stays still determines what can be seen.