Your own UI AI lab,
running locally.

Every AI product this year says "personalized." Ask any of them for the experiment behind the word and the room goes quiet. uistash takes the opposite bet: ship the measurement, let it argue.

There's a harder problem underneath, and the Lab is built around it: when you test your own taste, the experimenter and the subject are the same person — and the person most motivated to see the "with taste" arm win is you. So the Lab isn't designed to convince your followers. It's designed to survive you.

The lock is the only referee.

An experiment is a pre-registered bundle: the briefs, the arms, the number of rounds, and the stopping rule are all fixed before the first round generates — then sealed. No editing mid-run, no stopping early on a hot streak, and no running score anywhere until the series completes.

happiness-v1 · sealed experiment
briefs12 · locked
armshidden until complete
stopping rulefixed N · pre-registered
progress3 of 12 judged
running scorenot shown. ever.
Briefs and arms are sealed. Abandon and restart to change them.

This is friction on purpose — friction as scientific integrity. Convenience is what p-hacking feels like from the inside.

You're allowed to say neither.

Each round shows candidates blind — shuffled, unlabeled, identical chrome. Most eval tools force a winner. The Lab gives your eye its full vocabulary:

Pick A Pick B Both Neither Unsure
Neither is signal
"Neither is shippable" is a real judgment, counted separately — not thrown away, never forced into a fake pick. Some briefs exist precisely to earn it.
Unsure is honesty
Can't tell? Say so. Unsure rounds leave the rankings entirely and live on as an honesty counter — noise the statistics get to see instead of swallow.
Decision before rationale
You verdict first; the why-chips come after. Reasons invented before a choice contaminate it — the Lab never lets the story get ahead of the pick.

Some rounds are traps. On purpose. For you.

Blinding the labels isn't enough when you know a "with taste" arm exists somewhere — wanting it to win leaks into your picks. So inside a sealed experiment, the Lab hides the composition itself, and salts the series:

catch rounds · ~1 in 5
  • Both candidates are bare — no taste arm exists.
  • You can't tell a catch from a real round.
  • You can't favor what you can't find.
  • Your picks on catches measure your bias directly.
the anti-self-deception machinery
echo rounds · zero tokens
  • A pair you already judged returns later — sides swapped.
  • No new generation. Costs nothing.
  • Agree with yourself? That's your consistency ceiling.
  • Always pick the left one? That's your position bias, caught.
the instrument measures its own judge

No other eval setup we know of measures the evaluator. At N=1, it's the only honest option: before the Lab claims your taste file works, it tells you how reliable your own eye is.

Steer yes, brag no.

Results speak plain language, computed from exact odds — never dashboard confidence. Seven of ten sounds like victory; it's a coin flip's good day, and the Lab says so to your face:

resultchance does thisthe Lab says
7 of 10~1 in 6chance explains this
8 of 10~1 in 18a lean — not yet evidence
9 of 10~1 in 93hard to attribute to chance
15 of 20~1 in 48hard to attribute to chance

A percentage over four picks is disinformation with an axis — so below ten rounds you'll see "5 of 8", never "62.5%". The word "significant" doesn't appear in the product. Neither does a p-value.

Numbers this product will never show you.

A measurement tool is defined as much by what it declines to display. These are constitutional — no setting turns them on:

refuse-listconstitutional
human + machine verdicts, blendednever
percentages under 10 roundsnever
category slices under 10 picks per cellnever
p-values · "statistically significant"never
a pass granted by the machine judgenever
rankings under 5 comparisonsnever
mid-experiment winner calloutsnever

The machine earns its opinions.

After you pick, a machine judge — anchored to your actual saved references, never to "good design" — predicts what you chose. It starts with zero authority and buys each role with accuracy:

A rights ladder
Commentary is free. Anything more is earned at sustained agreement with your blind picks — and lost again the moment agreement drops. Certification only counts on control rounds, where the judge can't grade its own homework.
The mirror alarm
Every few runs, the judge also scores with a scrambled anchor — someone else's taste, essentially. If the wrong anchor agrees with you almost as well as yours does, the judge isn't seeing you; it's flattering you. Its earned roles freeze, and the Lab says why.
Separate ledgers, forever
Your verdicts and the machine's live in different books, and no aggregate ever mixes them. The judge is a calibration needle on the instrument — it is not, and never becomes, the instrument.

Three questions only this can answer.

Does my taste file work?
Same brief, same engine, with and without your tastestack — judged blind, across a sealed series. If the file is real, your own picks will say so. If they don't, that gets published too.
Which model already sees like me?
Every engine has a different eye, and none is graded on yours. The leaderboard ranks them by your blind picks — with "too early to rank" written where the data is thin. Then you spend your subscription money on evidence instead of vibes.
How reliable is my own eye?
Consistency ceiling, position bias, bias toward wanting the taste arm to win — measured from echo and catch rounds you never notice. The Lab's boldest output isn't a score for the AI. It's an honest portrait of the judge.

The evidence is being collected. Join the study.

The instrument is built — sealed experiments, controls, honest tiers, the ladder, all of it, running locally on the buyer's machine. What's pending is the thing everyone else skips: the results. They'll be published either way — a negative result is still a contribution, and pretending otherwise is how this industry got here.