The correction entry must sit next to the error
Aliases: in-place correction · candidates beside ink · handwriting candidates
What it is
Once recognition emits characters during writing, the gaze is still on that ink when a wrong glyph appears. The correction entry has to sit beside the error: tap that ink, see candidates, swap a character, without first parking the pen on a distant candidate bar or keyboard. That is spatial coupling inside the recognise-and-correct loop. After the result is committed as ordinary text and edited like typed text, that is post-editing, not this in-place fix.
Why it happens
During writing, fixation follows the nib, and the visual evidence of a wrong character is the ink just left. If candidates appear along the bottom or on another panel, both gaze and pen make a round trip: find the wrong token in a list, then return to the line for context. The trip breaks the motor plan for the next stroke, so people often finish the line and correct in a batch, and the “live” feedback of recognition goes unused. An in-place entry stacks candidates beside or just under the ink; the hit target is that character, the action is “tap the wrong glyph → pick the right one”, coincident with attention. Low confidence can expose candidates before a silent wrong commit, so people do not have to discover the error first. The entry must also handle segmentation errors: two ink blobs fused, or one character split. The in-place selection must grow and split, or the only fix is delete-and-rewrite. With the pen as pointer, candidates need size and spacing so the palm and nib do not hit a neighbour.
Studying it
Compare in-place candidates, a bottom candidate bar, and keyboard-only correction. Seed sentences with high-confusion glyphs (similar shapes, cursive).
Independent variables: entry location (beside ink / bottom bar / keyboard), whether the list follows the current character, whether segmentation can be split or merged in place. Dependent measures: time from noticing to fixing, gaze leaves from the ink, whether the next stroke resumes, missed fixes.
Eye tracking or coded video shows the round trip; final accuracy alone does not.
Where it stops holding
Candidates stacked on the ink can cover where the next stroke will go, especially writing left-to-right with a panel opening to the right. On a small screen the in-place panel may hide context; it needs to be brief or translucent. Whole-word errors span more than one glyph; the entry must be able to select the word, still at that word’s ink, not across the screen. An external keyboard as an accessibility path is a supplement, not a reason to remove in-place entry.
Applying it
- Tapping (or pressing) the ink of a wrong character lists candidates that follow that block, not a fixed bottom bar.
- Mark low-confidence results lightly so they can be fixed while the person is still there, rather than becoming silent wrong text.
- Allow expand and split on adjacent ink for segmentation errors, plus “rewrite this block” without clearing the line.
- How to check: bury a look-alike error mid-line and see whether people fix it in place before the next character. Count pen trips from that glyph to a distant bar. Move candidates to the bottom as a control; round trips should rise.