Notes · 03

The transfer test.

Every AI product this year says it is personalized. Almost none of them show you the measurement. A demo is the easiest thing in the world to fake: run the prompt ten times, keep the best screenshot, throw the other nine away. This note is about the instrument uistash uses instead, and about what it can and cannot say yet.

01 · Demos lie

A good output is not evidence.

A single good result proves that a good result is possible, which you already knew. It says nothing about whether the thing you added, a person's taste, is what made it good. Any prompt run enough times will hand you something worth screenshotting.

To learn what actually caused the quality, you hold everything else still, change one variable, and look at many runs instead of the flattering one. That is an experiment. The distance between a demo and an experiment is the distance between a claim and a fact.

02 · Change one thing

Same brief, same engine, three diets.

The transfer test runs one brief through one engine three times, and the only thing that differs between the runs is what each is allowed to read. The first gets the brief alone. The second gets the brief plus the original screenshot, which is how nearly every tool works today. The third gets the brief plus the coded reference exactly as an agent receives it over MCP: the tokens, the DESIGN.md, the key source files.

One aesthetics-silent brief feeds three lanes that differ only in what they carry: control gets the brief alone, image adds the screenshot, mcp adds the coded reference as tokens, DESIGN.md and source. All three outputs are scored against the original at the same canvas. Two deltas are named, not numbered: mcp minus control asks whether the reference transfers, mcp minus image asks whether it beats the screenshot. briefaestheticssilent briefcontrolnothing else briefscreenshotimagea picture brieftokensDESIGN.mdsourcemcpthe reference scored vsoriginalsame canvas mcp − controldoes the reference transfer?mcp − imagedoes it beat the screenshot?
Same brief, same engine. Only the information diet changes. The deltas are the whole argument, and they stay unnumbered until the study reports.

Same words, same model, three different information diets. Whatever separates the outputs came from the material, because the material is the only thing that changed.

03 · The silent brief

The brief says nothing about looks.

For the comparison to mean anything, the brief has to describe what to build without describing how it should look. It names the content and the structure and stays silent on aesthetics, so any style in the output had to arrive through the reference material, not through the instructions.

The brief is scrubbed against the reference's own values before it runs, so a hex color or a font name cannot leak in through the back door. If the mcp run looks more like the original, it is because the reference carried the look, not because the brief quietly described it.

04 · What survives

A picture and a reference are not the same payload.

This is the quiet claim underneath the whole product. Flatten a design into a JPEG and the pixels survive while the values fall away. The exact hue, the shadow, the easing curve, the behavior on click: none of it is in the file, so anything building from the picture has to imagine it, differently every time.

Two stacks of the same five design concerns. In a screenshot only the pixels survive and the hue, shadow value, easing and interaction are imagined. In the coded reference each concern is carried as a real value: a token hex, an exact shadow, a cubic-bezier, working code, and the reasoning in a DESIGN.md. a screenshothueimaginedshadow valueimaginedeasing curveimaginedinteractionimaginedpixelskeptA picture keeps the pixels. Everything else is guessed, differently every time. the referencehue→ token hexshadow value→ exact px/rgbaeasing curve→ cubic-bezierinteraction→ working codewhy→ DESIGN.mdThe extraction keeps the values, once, so an agent reads them instead of imagining them.
A screenshot and a reference carry different amounts of the original. The test measures the gap instead of asserting it.

The extraction keeps those values as text, once, so an agent reads them instead of guessing. The test exists to measure that gap rather than assert it, which is the difference between a slogan and a number.

05 · Two deltas

Not yet numbered, on purpose.

The experiment produces two differences. The mcp run minus the control run asks whether the reference transfers at all. The mcp run minus the image run asks whether the extraction beats the screenshot the rest of the world relies on. Both are scored the same way the import pipeline scores fidelity: against the original, at the same canvas, region by region.

honesty notenotes · 03
the instrumentbuilt
the briefswritten
published resultsnone yet
numbers on this pagezero
there are no numbers here because there are none to report yet

Inventing a result would break the only promise the method makes, which is that a claim has to be earned. When the study runs, the deltas will say what they say, in the open, whichever way they fall.