Repository navigation
Record what the 156 file survey actually said - #3
Merged
Merged
Conversation
The run that produced the 1.1.1 fixes died partway through with the session that started it, so it was finished and the numbers are now in. Zero errors across 156 files. The same 66 files had already been run against 1.1.0, which makes this a head to head rather than two samples. Exactly three files changed verdict, all field_blended to undetermined, and nothing else moved. Both blend cycle false positives are gone and no new ones appeared anywhere in the 156. The 71% undetermined rate needed unpacking rather than reporting. 38 of the files are one essay video's parallax stills, correctly undetermined because there is no cadence in a slideshow. Of the rest, 108 of 118 are short enough that the sampler gets one to three windows instead of twelve, and undetermined runs 76% on the one window files against 53% on the two to three window ones. Window count is the variable, not duration: I had assumed short clips were the problem and long ones were fine, and measured over the whole set it is the long files that do worse here, because the long files are the slideshows. Also recording an unresolved case rather than claiming it either way. ep01_nomusic.mp4 reports mixed with four field_blended windows, and those windows are not the dead band case just fixed: dirty is 1.0, field matching moves the comb figure by under 0.3 points, and 30 to 44% of frames solve as a blend. That is the signature the rule describes, and the material is parallax scrolled stills which really can blend. There is no ground truth for it. README gains the short file explanation next to the undetermined verdict, since the note carrying that information is currently buried under the window table. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
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
1 participant
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.
Finishing the loose end from #2. The survey run that produced those fixes died partway through, so it was restarted and completed: 156 files, zero errors.
The same 66 files had already gone through 1.1.0, so this is a head to head.
Three files changed, all
field_blendedtoundetermined(kohana_motion2,liqchrome_angel,liqchrome_dragon). Nothing else moved, which is what you want from a gate tightening rather than a threshold nudge. Both blend cycle false positives are gone (rin_hq_finalwas claiming a 52.5 fps source,rin_shift2a 21 fps one) and no new ones turned up in the other 90 files.The 71%
undeterminedrate across the full set looks bad and mostly is not. 38 files are one essay video's parallax stills, which are static slideshows with no cadence to find. Of the other 118, 108 are short enough that the sampler gets one to three windows instead of twelve, andundeterminedruns 76% on the single window files against 53% on the two to three window ones. Only 5 files in the set are long enough for the full 12 windows.I had said earlier in this work that
undeterminedwas inflated by clip length. Measured over the whole set that is the wrong variable: it is worse on the long files here, because the long files are the slideshows. Window count is what predicts it.One case is left open rather than resolved.
ep01_nomusic.mp4reportsmixedwith fourfield_blendedwindows, and those are not the dead band case #2 fixed. They sit atdirty=1.0,clean=0.0, comb 17 to 29% of blocks, field matching moves the comb figure by under 0.3 percentage points, and 30 to 44% of frames solve as a blend of their neighbours. That is exactly the signature the rule describes, and parallax scrolled stills in an editor really can produce inter-frame blending, so it may be right. No ground truth either way, so it is recorded.README gains a paragraph next to the
undeterminedverdict explaining the short file case, because right now that information is in a note under the window table where nobody reads it. Docs only, no code change, so PyPI's description stays a paragraph behind until the next release rather than earning a 1.1.2.122 tests pass, ruff clean.
🤖 Generated with Claude Code