Actionist · pack picker research · 27 Aug 2026

Ten Clicks to Taste

You can't ask a client what they like. You can watch them choose. The science of turning a handful of clicks into a design-preference profile — with the receipts, and a toy you can play.

Play it first

Below are four small interfaces. Pick the one you'd want your product to feel like. Then do it again. The panel underneath is the system's live belief about your taste — watch it sharpen as you click. If a round is all wrong, re-roll: that's a signal too.

The toy uses win-rate counting with Laplace smoothing — a zeroth-order Bradley–Terry. Production would fit the real model (see choix) or a Gaussian-process preference model. The logic is the same; only the statistics get sharper.

What just happened, from first principles

1 · Choices are data; statements aren't. People have no reliable introspective access to their design preferences — ask "warm or cool neutrals?" and you get noise; show two rendered screens and the answer is fast and consistent. Economics calls this revealed vs stated preference, and conjoint analysis built an industry on it: never ask about attributes, make people choose between bundles of attributes, and the attribute weights fall out of the choice pattern mathematically. knowledge — verify before client use

2 · Each card is a bundle, so each click is many signals. Every specimen above is a vector of five knobs (palette, radius, type, density, shadow — the demo's subset of our seven-knob pack schema). When you pick card A over B, C, D, the system doesn't learn "likes A." It learns A's palette beat three other palettes, A's radius beat the radii it was shown against, and so on — one click, five comparisons. That's why it converges in ~10 picks instead of ~50.

3 · The math has a name. Modelling "picked one out of a set" is Luce's choice axiom; the pairwise special case is Bradley–Terry, the same family as chess Elo. choix (MIT-licensed Python) implements exactly our observation type — top-1 choice from a shown set — and infers latent item strengths with proper uncertainty. verified 27 Aug

4 · Which four to show next is the clever part. The demo picks its next round by finding the least-resolved knob and building cards that disagree on it while holding your likely preferences on the rest. That is the toy version of preferential Bayesian optimization: model the user's hidden utility function, then query where information gain is highest. The direct prior art is Sequential Gallery (SIGGRAPH 2020, MIT, AIST): show a grid of design variations, user picks their favourite, the model updates, the next grid is chosen to maximise learning — iterate until converged. That is, almost word for word, the loop you just played. Its ancestor Sequential Line Search (SIGGRAPH 2017) did the same with sliders. verified 27 Aug

5 · "None of these" is a measurement, not a failure. Discrete-choice models call it the outside option — the choice to walk away — and it carries real information: weak negative evidence against everything shown, and a guard against the forced-choice bias where a cornered user picks the least-bad card and pollutes the data. The demo logs re-rolls as soft evidence and re-explores. knowledge — standard econometrics

Why this scales to unlimited packs

Learn in knob space, not pack space.

This is the load-bearing design decision. If the system learned "which of N packs do they like," every new pack would dilute the data and lengthen the quiz. Instead it learns a profile over the seven mechanical knobs — a fixed, small space that never grows. The pack catalogue can go from 16 to 16,000: the elicitation still converges in the same ~10 clicks, and the converged profile becomes a nearest-neighbour query over whatever catalogue exists that day. New packs are indexed by their knob vector the moment they pass the gates; the quiz never changes.

The same profile transfers downstream: it can pre-filter which heroes and sign-ins the component picker surfaces (compact-density people never see spacious editorial heroes first), and it can seed the image-gen design rounds. One elicitation, spent three times.

The supporting evidence for the gallery around it

From this project's verified research reports (all in SISO/research/, dated 27 Aug): choice overload is real specifically when preference uncertainty is high and options are hard to compare (Chernev et al. 2015 meta-analysis, N=7,202) — which is why every card renders the same reference screen; simultaneous presentation beats sequential on decision quality and commitment (Basu & Savani) — which is why it's a grid, not a wizard; and differences between options must be explicit, not implied (NN/g) — which is why cards carry their knob tags. verified in ui-pack-gallery-ux report

Three classics I'm citing from training knowledge, flagged honestly, worth a verification pass before any client deck: showing users multiple design alternatives elicits genuine critique where a single design elicits polite approval (Tohidi et al., CHI 2006); parallel variant exploration measurably beats serial refinement on real outcomes (Dow et al., 2010); and visual-appeal judgments form in ~50ms and are stable (Lindgaard et al., 2006) — which is why judging small rendered thumbnails is legitimate at all. knowledge — verify before client use

Sources

Written for the Actionist pack-picker workstream. The page itself is dressed in the Paper & Ink pack authored and gate-checked this session — every chrome colour above is one of its tokens.