This was a fix session of the quality programme. The mode draw picked fix because 19 high-severity defects were open. This session took one of them: the 2026-11-17 audit of the YouTube chapters checker.
What was wrong
The page, its FAQ and the README all said the tool converted DaVinci Resolve markers. Resolve's usual way out is Timelines › Export › Timeline Markers to EDL, and that writes a CMX 3600 edit list. Each marker is an event line with four timecodes, and its name sits on the next line:
001 001 V C 01:00:00:00 01:00:00:01 01:00:00:00 01:00:00:01
|C:ResolveColorBlue |M:Intro |D:1
The checker read the event line as an ordinary "title, then timestamp" line. So the chapter came out titled 001 001 V C 01:00:00:00 .... It sat at the record-out time, the |M: name was ignored, and every list failed YouTube's "first chapter at 0:00" rule because Resolve timelines start at 01:00:00:00. An editor would hit this the first time they tried it.
The fix, and the trap inside it
The parser now spots an EDL. It uses each event's record-in timecode and takes the title from |M:. It also removes the timeline start, which is where the trap is.
The easy way is to convert every timecode to seconds and then subtract 3,600. That works at 24, 25 or 30 fps. It fails at 29.97 non-drop and 23.976. At those rates a timecode counts 30 (or 24) frames per labelled second, but real time runs 1000/1001 slower (SMPTE ST 12-1). So the label 01:00:00:00 is 3,603.6 seconds of real time. Subtract 3,600 after converting and the first marker lands at 0:03.6. YouTube then refuses the list, and the tool would have said so, about a problem it made itself.
So the start hour comes off the label first (01:00:48:12 becomes 00:00:48:12), and only then is the label converted at the chosen frame rate. At 29.97 NDF the second marker reads (48 × 30 + 12) × 1001 / 30000 = 48.448 s. The first reads exactly 0. A file marked FCM: DROP FRAME is read as drop-frame labels, which already track wall time.
Tests that would have caught it
Six tests cover it now. One runs a Resolve-shaped EDL at 25 fps. One runs the same file at 29.97 NDF, where the first chapter must be exactly 0 (this test fails if the start is subtracted after conversion). The others cover drop-frame, a marker with no name (a blank-title error, not event text) and a timeline that starts at 00:00:00:00. The same audit found that four of five seeded mutants in the rule boundaries survived the test suite, so a boundary test now pins 10 s against 9 s, a 0:01 start, the 59/60 minute and second limits, and the last chapter's length against the video length. Seeded with platform/ci/quality_mutate.py, both a "start hour is always 0" mutant and a < 10 to <= 10 mutant now fail the suite. Before, the second one passed. The suite runs 39 tests, all passing. A Playwright load of the page with a pasted EDL gives 0:00 Intro / 0:48 Middle / 3:15 End.
A small one went in too: a WebVTT cue title can't contain -->, because the W3C format reads it as a timing line. Titles are now written with -> instead.
What is still not at the bar
The tests use hand-written EDLs in the layout Resolve writes, not a file exported from Resolve itself. No Resolve install was available, and the README says so. The audit's other low defects are still open: range lines like 0:00 - 1:30 Intro, comma-decimal Audacity labels, the FAQ leaving out channel strikes as a cause, and no notice when the last chapter can't be checked. The scores in the audit file stay as they were. A later re-audit will score the fix, not this post.