Matt Stokes · Design Exploration

Adaptive Density

A high-level UI paradigm in which visual density reflects how settled the system's current interpretation is.

AI interfaces usually render a result after the model has already collapsed uncertainty into an answer. The same visual finish can hide the difference between strong support and a guess.

Adaptive density is a high-level UI paradigm for keeping that uncertainty available to the interaction. Density rises as evidence converges and remains coarse while an alternative interpretation could still change the action.

The figures are system schematics. They show how evidence, alternatives, and state drive the paradigm. Screen-level expression remains open.

Rise and settle are system transitions

The system can increase resolution once the evidence supports a useful commitment. Attention moving elsewhere does not settle the state. An unresolved interpretation should stay coarse until new evidence either supports or removes it.

Rise & Settle
Interactive figureOpen
Fig 01 · Evidence rise and settlement · system schematic

Evidence layers stay separate

Support rarely comes from one signal. Spatial proximity, timing, semantic relevance, and interaction history operate on different timescales.

Combining those signals into one score too early hides disagreement. The system should retain the separate weights until the conflict can no longer change what it does.

Context Layers
Interactive figureOpen
Fig 02 · Evidence layers · system schematic

Alternatives remain active while they can change the action

The system may have several plausible readings of what a person wants. New evidence can change their order quickly. Removing every alternative makes the leading answer look final before the evidence supports it.

Any alternative that could still change the action remains active in the system model. The diagram shows that inference process rather than a proposed set of on-screen choices.

Intent Inference
Interactive figureOpen
Fig 03 · Intent inference · system schematic

System selection changes interface authority

Most applications make people choose where to work and gather the relevant context themselves. An ambient system performs part of that selection before anything appears.

Every selected item carries a judgment about relevance. Once the system controls that selection, its evidence state becomes part of its authority. Before an ambient system files a message or books a room, the product should preserve why the action was selected and how much uncertainty remains.

Selection Shift
Interactive figureOpen
Fig 04 · Selection shift · system schematic

System state controls visual resolution

The system-state model tracks where evidence has converged and where it remains open. Adaptive density uses that state to govern visual resolution.

The diagram uses a monochrome field to explain the mechanism. It does not prescribe stems, peaks, fields, or any other screen component.

System State
Interactive figureOpen
Fig 05 · System state · system schematic

Belief updates remain continuous

Each observation changes the support assigned to the current interpretations. One area can become more settled while another remains open.

A hard cutoff collapses that range too early. The belief model keeps the distribution active until the remaining alternatives can no longer change the action.

Belief Engine
Interactive figureOpen
Fig 06 · Belief update model · system schematic

Competing interpretations remain available

The leading interpretation receives more weight inside the model. A plausible alternative continues gathering evidence and disappears only after the system has ruled it out.

The eventual interface can express that state through hierarchy, detail, pacing, or another treatment appropriate to the product. The figure shows the internal logic that any treatment would need to preserve.

Decision Fabric
Interactive figureOpen
Fig 07 · Competing interpretations · system schematic

Evaluate the paradigm through correction

Many models already estimate uncertainty. Product interfaces often hide it before presenting the result.

Adaptive density keeps the evidence state available to the interface without prescribing one visual form. A writing tool could keep a weak suggestion coarse until its premise is accepted. An agent workflow could increase detail before the system saves information or commits an action.

The first product test should measure whether people can distinguish a settled interpretation from an open one and correct the system before an action commits.