A working sensor tells you what is on the road. It can't tell you who is behind the wheel at hour eleven, what a fleet is legally on the hook for when an alert gets missed, or whether the thing you built survives contact with a real driver. Those answers don't come from the hardware. They come from research: crash data, human-factors studies, the regulations that govern how long a person can legally drive. That work is what separates a design decision from a guess, and every choice in this project traces back to it. Three lenses, in the order I used them: the driver, the business, and the proof.

Ray isn't a novice to be protected from the machine. He is an expert whose attention is a finite resource, spent down over a regulated eleven-hour window. The design has to account for a professional whose skill is real and whose alertness is not guaranteed.

Here is the insight that reframes everything. The better the automation performs, the worse the human gets at stepping back in. Skill you don't use decays. So the interface can't just work well; it has to keep Ray in the loop precisely because the system is good enough to let him drift out of it.

And the record lies about him. Fatigue barely registers in the official crash-cause data; inside real cabs it is everywhere the paperwork isn't. Design for the Ray the paperwork describes and you are designing for a driver who does not exist.

Grounded in Bainbridge, "Ironies of Automation" (1983); FMCSA Hours-of-Service rules; naturalistic fatigue prevalence from VTTI and FMCSA studies.

A fleet doesn't buy a driver interface to be elegant. It buys down risk. Every decision on the screen sits inside a knot of forces that pull against each other, and the obvious solution usually serves one while quietly wrecking another.

Suppress every alert to keep drivers happy and you raise liability. Alarm on everything to cover the fleet legally and drivers mute the system, which is worse than no system at all. The design lives in the tension, not on either side of it.

And the reason this is worth money is measured, not assumed. On heavy trucks, forward collision warning and automatic emergency braking cut the specific rear-end crashes they target by 44% and 41%. That is the business case, in someone else's data, before I draw a single frame.

Framed against SAE J3016 (the industry's self-driving levels), IIHS effectiveness studies on heavy-truck FCW and AEB, and the NHTSA heavy-vehicle AEB rulemaking.

Everything above this line is built on published research. Federal crash data, naturalistic driving studies, human-factors literature, regulatory filings. That work is real, and it is enough to frame the problem and shape a first design. It is not enough to ship one.

Secondary research tells me what happened to drivers in aggregate. It cannot tell me what this fleet's safety director will tolerate in liability, what a dispatcher's schedule does to one route, or what Ray does with his hands in the ninety seconds after a false alarm. UX does not live in the aggregate. It lives in the person. Here is where secondary research ends, primary research begins, and what each method settles.

The image is the output. The prompt is the system behind it, and I wrote it as a design specification, not a loose visual description. Every section does a job. Role sets the kind of image. Scene fixes the context. Governing Principle protects the hierarchy. Focal Hazard names the one thing that earns attention. Constraints keep the model from adding noise.

The load-bearing decision is restraint. In a safety interface, every lit element competes for the driver's attention, so amber is rationed to the stopped vehicle and its alerts. One hazard is easy to read. A dashboard full of warnings is not. The phrase "You decide" does the same work in words: the system informs and supports, the driver still owns the action.

Image models struggle with small interface text, so the rendered labels are placement and intent, not final copy. I'd set the real UI text in Figma, where accessibility, content, and interaction are decisions, not a render. Concept exploration is one job. Human judgment on the details is another.