Repository navigation
fix(notes): find a one-word quote inside a word only from three letters - #666
Merged
Merged
Conversation
… "and" A drag from inside one speaker's line into an overlapping line of another is quoted in time order (Valtora#616), so the second line's first words come before the first line's last. No order of whole lines puts them next to each other, so the resolver marked unchanged words as changed, on partly wrong words. Where lines overlap in time and are drawn side by side, a match may now move from one line to another as such a drag does: inside a group of overlapping lines from any word to any word, and into the next group from a line's last word to a line's first. A move is a tie-break, like a part-word match, and each line's part may start or end inside a word, as the quote's ends may. None of the 10,500 calibration selections crosses lanes, and every one resolves as before. A one-word quote matched inside any longer word, so "a" landed on "and" as an exact match once the original "a" was gone. It now matches inside a word only when it is three letters or more; a whole word still wins. On the calibration sets (2 s, 0.5), wrong marks fall from 124 to 119 of the 7,500 selections and from 39 to 18 of the 3,000 drags, and 105 drags that took one or two letters of a word are now unanchored. Two letters would keep 46 of those and leave 10 more wrong marks; four would lose 30 more right ones for 7 fewer wrong. Most of the wrong marks left (92 of 119) are one-word quotes found whole on the same word said again within 2 s, which no length rule reaches. The property tests' seeded generator overflowed a double's 53 bits and repeated after about 10,000 draws, a few dozen distinct cases per test. It is now mulberry32. With it, the test that never marks words once the quoted ones are gone fails on the old resolver ("a" inside another word). The identity test now meets a selection that starts on the punctuation after a word the same as the one picked, which a one-word quote can't tell apart, and accepts either. Addresses Valtora#488.
…e hold to let go With distinct cases the identity property resolves a 2,000-character quote in a third of them, and took 5.3 s on CI against its 5 s limit (about 2 s here; the resolver's share is 1.6 s against 1.3 s before the line moves). It now runs 1,000 cases with a 20 s limit, its floors scaled to match. The editor releases the page in its unmount effect, which React may run after the commit that removes the dialog, so the leave check after a save could still see the hold. It now waits for the hold to go, as it waited for the dialog; it failed 1 run in 20 under load before, none after.
Contributor
Author
|
Update on the two follow-ups: we're keeping the one-word rule and dropping the overlap-lane change from this PR. Letting the match step between overlapping lines fixed your drag, but our review found it also turns a reworded single-line quote into an "exact" match by borrowing a word from the other speaker's line (about 40% of probe cases inside an overlap group, against under 2% before), and it is slow in crosstalk (4 s against 28 ms for a long quote with 20 overlapping speakers). Constraining it to drag-shaped matches works but needs a rework of the search for the speed. Overlapping lines are rare in finished transcripts (none in the 4,933 lines our calibration resolves against), so we'd rather leave the lane case unhandled until someone hits it than add that machinery now. The PR is the three-letter rule and the test generator fix. — nullcat (Claudette) |
The moves between lines drawn side by side found a drag across two speakers' lines, but they also let a single-line quote with one word changed borrow that word from an overlapping line and read as exact (about 40% of probe cases inside an overlap group, against under 2% before), and they were slow in crosstalk: 4 s against 28 ms for a 350-word quote with 20 speakers in one group. Holding them to drag-shaped matches needs a rework of the search for the speed, and overlapping lines are rare in finished transcripts (none of the 4,933 lines the calibration resolves against), so the lane case is left until someone meets it. The resolver is upstream's again, with the three-letter minimum for a one-word quote; its calibration figures are unchanged (the moves never fired there). `selectionRange.ts` is upstream's, and the lane tests and the ARCHITECTURE sentence go. A new test keeps a one-line quote with a word changed fuzzy on its own line while another speaker's line beside it says that word; it passes on the old resolver and failed with the moves. The identity property's exemption for a selection started on a twin word's punctuation now requires that the selection took only that punctuation.
thebeline
changed the title
thebeline
marked this pull request as ready for review
October 23, 2026 19:42
Valtora
approved these changes
Oct 23, 2026
Valtora
left a comment
Valtora
left a comment
Owner
There was a problem hiding this comment.
Approved. The three-letter floor only touches one-word part-word matches, a whole word still wins, and the mulberry32 generator fixes the repeating draws.
5 of 14 tasks
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Pull Request
Description
Addresses #488: the one-word follow-up from #659. The overlap-lane change is left out (see thread).
Type of change
Checks run
Migration impact
Documentation impact
docs/ARCHITECTURE.md(highlight rule).Security impact
Manual verification
Three new tests fail on
main; all 6 mutants fail one. Resolver only; no browser pass.