The model can't roll its own dice

We asked for design concepts under sixteen different creative framings: a bland-design-epidemic pep talk, an impossible-to-please client, taste vocabulary, world-building metaphors. Thirty of thirty-five answers were the identical concept. Temperature buys you different sentences, not different ideas.

Exhortation prose changes the skin and never touches the concept. This is the finding that kills most people's first instinct.

"Write down your first three ideas, discard them unseen, build the fourth." It works: the model escapes its #1 concept, and lands on its #2 concept. Every time. Across every rejection-shaped operation we tried, including cross-pollination where the model picks its own adjacent field.

The concept prior for a given category is about two ideas deep. Telling a model to avoid X gets you everyone's second idea, which is a different rut, not creativity.

Forcing a real designer's procedure, deriving seven candidate directions from the audience's world with a written rationale for each, produces genuinely diverse lists. Then "build the most resonant one" crowns the same winner in every run. Diversity exists in the model's candidates and dies at its selection step.

The fix: the model derives the grounded shortlist; a script rolls which index gets built. The dice never touch an ungrounded idea. They only refuse the argmax rut. A later A/B settled which word carries the weight: hand the model a shortlist and let anything pick from it, itself, a veto floor, a simulated user, and every chooser reverts to argmax, option one in 27 of 30 packets. So the dice have toassign the index; the moment they merely nominate a menu, a taste function collapses it again. And the script prints a reproduction key with every roll, so a bad draw in the field is a replayable bug report, not a mystery.

For a product rooted in Polish TV culture, the model's own candidate list was wonderful seven deep: the end-of-film voice credit, the newspaper TV guide, teletext at #3, the video-rental shop, the dubbing script. For SRE tooling, seven-deep is still terminals and postmortems. A monoculture home yields a monoculture list.

That's when assigned foreign forms earn their keep: a naturalist's field guide, a broadsheet sports section, injected from a curated pool and weighed against the derived list on two axes. Does the audience identify with it; does it make the product clearer. They lose to strong cultural material and win over thin categories. Which is exactly the shape you want.

We shipped a "costume check": if a visitor would notice the borrowed form before the product, the candidate fails. Sounds wise. Then we noticed that our 10/10 human reference, a real site built entirely as teletext, fails that check. You notice the form instantly; it works anyway, because the task stays easy.

Count the brakes in your prompt against the accelerators. Ours ran five to one, and the model does that arithmetic: it lands at 65% commitment, and 65% commitment is what "bland" looks like. We deleted the check and inverted the order: land fully committed first, that's the hard part, then make it clear and effective in later passes. Biggest single quality jump of the campaign.

Concept, palette, composition, motion: each converges independently, and dislodging one leaves the others at default. Our best failure was a page whose every atom was committed teletext, page-number chips, dithered level bars, the four color-key buttons as the nav, sitting on a completely standard marketing grid. Committed skin hides template bones from any judge that looks at styled output, including the model judging itself.

The diagnosis held; our first two remedies did not. We asked the model to strip color and texture in its head and check the bare wireframe against the category's standard page. Told to do it silently, it wrote no verdict and shipped the template anyway, pure audit theater. Told to write the verdict out, the ceremony taxed the build and quality dropped, so we cut the build-time check entirely. What stands is one generative line,borrow the form's skeleton, not its clothes: let the form set the structure, and a committed skin on a template grid is the tell it only reached the surface. Skin-blind judgment survives too, but as a review instrument, a wireframe grader and a fresh reviewer at the finish, never as the builder grading itself mid-flight.

Contracts promised radical structure ("the page IS a distributed trace; section headers are span names with durations") and the build shipped a default hero anyway. Nothing ever compared pixels to promise. Generic quality critics made it worse: pages got busier, not better, because a rubric invites additions.

What worked: the model writes its direction contract into the artifact itself as a comment block, five short blocks at the top of the file. We first fed it back inside the same build thread at completion, and a thread grading its own work rubber-stamps it, the same theater lesson six ran into. So the audit moved out: aseparate reviewer agent, a fresh reader, checks the render against the contract promise by promise and rebuilds the gaps, while the build thread stays building. Self-accountability has ground truth; rubrics don't.