01 — Three dimensions
Task, stimulus and time define the observation situation.
Searching for a target directs attention differently from free viewing. A webpage imposes vertical structure, text and scrolling absent from a natural image. The first seconds often establish global organization before more local processing.
These dimensions directly change fixation duration, saccade amplitude, dispersion, mode switches and mouse interaction. The protocol therefore treats them as experimental factors or model variables.
02 — An empirical example
The same participant behaves differently across situations.
In our Scientific Reports study, 91 participants viewed images and webpages freely or while searching for a target. One-second analyses showed significant task, stimulus and time effects on nearly every metric considered.
Webpages produced more vertical exploration and more mode switches; target search changed dispersion and depth of processing. Averaging those conditions would have created an artificially smooth portrait.
03 — Robustness
Test effect stability across several conditions.
To call something a personality signature, we looked for relationships that persisted for several seconds and kept the same direction across contexts. This requirement reduced the number of retained effects while increasing their theoretical value.
A condition-specific effect may reveal an interaction: a trait is expressed when the task offers particular freedom, or when a stimulus demands a specific strategy.
04 — Modeling
Use context as a variable, a layer and a test.
One strategy includes context in a hierarchical model: measurements nested in trials, nested in people. Another trains in one context and tests in another to quantify generalization. Performance can also be reported by sub-condition instead of as a single global average.
For applied systems, the final question is explicit: for which population, interface, task and time window is the result valid? That sentence belongs to the model as much as its coefficients do.
05 — From concept to decision
The test split should reflect the product promise.
A random trial split can place one person’s traces in both training and testing. I would therefore split by participant to assess new-user use, and by website to assess use on new interfaces.
These tests answer different questions. Context-specific performance and uncertainty show where a model is usable and where more data or adaptation are needed.