These are not pictures. All three are the shipped code — card, tokens, wordmark and every row read out of the built widget — and they work: drag a rubric row by its grip or use the arrows, tap a side, type your own phrase, then finalize and watch it save.
Each one resets underneath, so you can run a state more than once. The one worth sitting in is Saving: tap Finalize and watch the pause before it settles.
The order is the whole control. There is no second section and no tier to assign — you put these in the order that matters and finalize. An earlier cut split the list into what the interview asks about and what it only reads, which added a rule you had to learn before the control made sense.
What Claude picks, and what it does not. Claude sends the criteria and the topics under each one; the ordering is yours. It does not get to reword the parts that have to mean the same thing across two different rubrics, or they stop being worth comparing.
Both ways move a row. Drag by the grip, or the arrows — and the arrows are arrow keys too, so none of this needs a pointer. The buttons are the point; the grip and the hover lift are there so a row looks like something you can pick up before you try.
One thing worth deciding. Tap Finalize and watch the pause. The only thing that changes while it saves is the act dimming — nothing says work is in progress, so a slow save reads as a dead button. A label change or a spinner would fix it; that changes shipped behaviour rather than styling, so it is a decision rather than something this pass made.
Nothing here is merged. The connector's rubric still renders the old importance sliders. This control is built and not yet wired in.