Skip to content
JAEYONGBAE
EN/KR

All experiments

When a model assembles the screen at runtime, what to hand it and what to keep.
AI, 2026

Generative UI

The question

When a model builds the screen at runtime, what does it get to decide and what does the system keep?

Three projects settled it into one sentence. Content from the model. Form from the system.

Why the line is needed

Let a model build the whole screen and consistency and trust go first. Fix everything in advance and there is no reason to use a model at all. The usable answer sits between, and the three projects put the line in three different places.

Evidence

  • The model proposes, a person decides – SAI – the suggestion card is a fixed component with four states (offered, working, done, dismissed). Only what is inside it is generated. Only what a person accepts becomes a note, and suggestions never mix into the note itself. At most two wait for an answer at once (design value).
  • No evidence, no generation on screen – SAI – a suggestion without a source utterance or material is not shown. AI output carries a blue glow and an "AI-generated" badge, which is also how Korea's AI Basic Act disclosure duty is met. Confidence shows as the fill of the card, and what it doesn't know it says it doesn't know. New material queues behind a banner instead of replacing the card being read.
  • No runtime model call at all – Commit – grant summaries and eligibility checks are rule-based. There are no runtime LLM calls, for two reasons: cost and hallucination. Amounts, dates and eligibility come from normalised data, and where the rules are unsure the figure stays empty. 2,595 of 11,757 notices carry an amount, 22.1% (output).
  • A fixed form receives the model's probabilities – Distance of Thinking – the six next-token candidates the model actually produced, and their probabilities, are built as a space. Generated text does not make the screen; a fixed form receives the model's computation. No colour is used: meaning comes from density, opacity, size and movement.

What holds

Form has to stay put for what someone learned to keep working. If the component changes every time, the user learns it every time.

Generated output without evidence is not something to accept. People decide to press once they can see what it came from.

Runtime generation is a cost. Whatever could be computed in advance was. In Commit, not calling a model at all was the right answer: not creating the place where it can be wrong was cheaper than correcting wrong answers.

Not verified yet

  • Acceptance rate – which SAI suggestions get accepted and which get ignored is not measured. Without that number there is no basis for changing when a suggestion appears.
  • The two-suggestion limit – a design value. Nobody counted how many are right in a real meeting.
  • The limit of rules – of the amounts Commit's rules left blank, it is not clear how many are genuinely unknowable and how many would be recovered by writing more rules.