Entry patterns & hypotheses
Both conditions share one classifier, sample prompts, and results, so differences come from the entry experience alone. A real visitor would see one design; the review controls on the Ask page exist only for this study.
Calm canvas (chosen direction)
A faint one-row Ask card with the call to action on the same row and a rotating spectrum ring on focus. When the first interpretation arrives, the card stretches downward and the card eases open and a calm two-line recognized strip settles into place; later interpretations update in place. “Adjust details” expands that same strip into refinement controls, on request only.
Hypothesis: Showing exactly what was understood, quietly, signals capability and builds confidence; offering refinement on request avoids turning the question into a form.
Risk: Refinement is easy to miss, so editing may be rare; a two-line summary can truncate on small screens.
Control · Static search bar
A conventional keyword search field with no live feedback. Same submit path and results.
Hypothesis: Reference condition for first-impression and task-time comparisons.
Risk: None. It is the baseline.
After you ask: two flows
Both render the same blended result, so the comparison isolates where the answer appears. How the question was phrased only changes the order: a lookup (a title, form number, ticker or record id) leads with that item and other matches; a natural-language question leads with the shaped answer, with related records below. Neither half is ever hidden, so a wrong guess costs a scroll, not a dead end.
Next screen, continuous
Ask navigates to a result page. The same Ask card comes along at the top, pre-filled and editable, with the blended result below. Where supported, a View Transition moves the card from the hero to the top.
For: Real navigation: Back, refresh and sharing behave natively; the no-JS path is the same page; a new heading is easy for screen readers.
Against: A page load between asking and answering, however smooth.
Same screen
The answer appears under the Ask card on the homepage; the headline collapses and the card docks at the top, in place. Back returns to composing with the text intact. Without JS it is one reload to the same page.
For: The most continuous feel: the question and its answer stay one object.
Against: Needs explicit history and focus handling (focus moves to the results heading on submit, the one exception to “never move focus”).
Explored and set aside
- Adaptive canvas. Understanding grew inside the field as a full panel. It was informative but visually busy, and the field kept changing size.
- Context sidecar. Understanding sat in a separate column. Eyes had to travel, and on mobile it collapsed into something like the adaptive canvas.
How to read the interface
- Recognized
- Recognized intent: strong, explicit evidence. Only now do editable controls appear.
- Tentative
- Some evidence. Gentle wording, no controls, so nothing jumps while you are mid-sentence.
- Topic: Clean energy
- Inferred detail: read from your words, display only. Dashed edge, labelled “from your words”.
- Set by you
- Explicit override: you changed a control. Solid edge, labelled, with a reset. Your text is never rewritten.
- ? Not in Atlas data: Apple
- Unrecognized or ambiguous: named but unknown, or several plausible readings. Shown, not hidden.
Interaction rules
- The composer never moves, loses focus, or changes text because of an interpretation.
- Every keystroke is classified locally (well under a millisecond). The visible surface only updates after a ~450 ms pause.
- If the words behind the visible interpretation disappear, its controls clear immediately instead of waiting.
- A stable interpretation stays put while you edit within the same intent (hysteresis), so it doesn’t flicker.
- Screen readers hear a polite, throttled summary only when an interpretation settles, never per keystroke.
- Asking is always one step. A suggestion never blocks or redirects submission.