# Cutlist Desk

> Deleting speech is subtractive, and subtractive edits fail in ways additive ones do not. This
> rebuilds the transcript in playback order, tests every repeat group against the rule that the
> last complete take is the one to keep, and checks the joins. Five lanes over one transcript and
> its cut list, plus a free browser-side checker.

URL: https://cutlist-desk.skillsafe.ai/
API: https://cutlist-desk.skillsafe.ai/api.html

Derived from Agentchengfeng/chengfeng-videocut-skills
(https://github.com/Agentchengfeng/chengfeng-videocut-skills), Apache-2.0: its semantic-deletion
reference, whose hardest lessons are about the input rather than the criteria and are both
measured. Not affiliated with or endorsed by Agentchengfeng.

## What the free checker does

It runs entirely in the browser, on the sheet you type, with no account and no credits.

1. **Rebuilds the stream.** Applies the cuts and produces what will actually play, in playback
   order, because adjacency after cuts is what every judgement depends on and it is not the
   adjacency in the file.
2. **Finds the repeat groups.** Segments that are takes of the same line, by shared opening,
   containment or bigram similarity - so a take that diverges at the third character is still
   recognised as the same line.
3. **Tests the ordering rule.** If what survives a group is not its last and most complete take,
   the rule has been inverted. This is the check the app exists for.
4. **Checks the joins**: mid-sentence splices, back-references whose antecedent is gone with
   nothing left in its place, and a piece that now ends on a connective.
5. **Checks the discipline**: risk labels against the shapes that are high risk whatever they are
   labelled, the two-state rule, stable ids rather than positions, silence left to the pause pass,
   and a submission smaller than the one before it.

## The ordering rule

Delete the earlier attempt, keep the later one. The last take is the one where the speaker finally
got it out. Almost every "why does this sound wrong" in a rough cut is this rule inverted: every
individual deletion looks defensible, and what plays is the attempt rather than the sentence.

The worst form is keeping a fragment - the complete line deleted, a broken opening left behind - so
what survives reads as a sentence that stops.

## The two measured input failures

- Keeping the original segment numbers. After earlier edits they skip, and a reader takes the gap
  as content withheld from them. Two correct calls were vetoed this way. A gap in the numbering is
  the trace of an edit, not hidden material.
- Marking compressed pauses as removed speech. The markers went from 15 to 58, the piece read as
  fractured, and the judgement turned uniformly conservative.

In both cases the criteria were right and the input lied, silently.

## Submission replaces, it does not accumulate

The set is rebuilt from the submission alone, so it must carry every round so far. Sending only the
new finds took a cut count from 430 to 394 upstream: the audio stayed correct because a manual
timeline only ever gets dug, but the record regressed and the call still reported success.

## What it cannot do

It cannot tell you which of two spellings of a name is right - only that the takes disagree and
that deleting words will not fix it. And where the string measures cannot separate a misspeak
retake from a take that continues into new material, it says so and asks you to read both rather
than guessing. That case errs toward asking: a needless glance costs a second, a missed one deletes
something nobody said twice.

## The lanes

- `read` - applies the cuts and reads back what will actually play, including the joins
- `check` - the ordering rule, the survivors, and whether each risk label matches its shape
- `round` - one scan pass at a time, carrying only that round's criteria
- `submit` - the whole conclusion, and what an increment would silently un-cut
- `repair` - turns the inversions back round

## What it does not do

It does not transcribe, cut or render anything. It takes a transcript and a list you already have
and checks what the two of them produce together.