Review the cut on the timeline

What you see when a job finishes

The job screen shows a Your rough cut is ready banner with an Export for Final Cut Pro… button, and beneath it the review list — one row for every decision RoughCut made about your footage.

Reviewing is optional. Exporting straight from the banner is a reasonable workflow for a clean recording. The list exists so you can check the editor's judgement before the cut reaches Final Cut Pro, and reverse anything you disagree with.

Nothing in this list touches your source file. Every row is a decision about the exported timeline only.

Timeline mode, and what it hides

A segmented picker above the list offers Timeline and By category. Timeline is the default. It sorts every item chronologically by start time, so reading top to bottom follows your recording.

It deliberately hides two kinds of row:

• items whose category is silence — the pauses that silence detection removed on its own
• micro-islands, very short surviving fragments that were merged into an adjacent cut

The section header states this plainly: Chronological — silences removed automatically are not shown.

That is a display filter, not a decision. Those rows still exist and are still reversible — they are listed in By category, under Removed: Silence and Removed: Other.

What each row tells you

Removed rows render struck through in a secondary colour. Kept rows render normally. A coloured bar on the left of each row keys the row to its category.

The reasons you will see

If you ran Second Look, rows it removed are prefixed Second Look — , and its borderline calls appear as kept rows reading Second Look suggests removing (…) — kept until you decide.

Confidence, and what the colour means

Confidence renders as an integer percentage pill on the row.

• Below 70% the pill is red, with the tooltip Low confidence — worth a look. Treat these as the rows most worth your attention.
• At 70% or above the pill is green, with the tooltip Editor confidence.

Some rows have no pill at all. Kept segments generated between removals carry no confidence score, because no model judged them — they are simply what is left.

A confidence figure is the editor's own estimate of how sure it is, not a measurement of correctness. A confident wrong call is possible. That is exactly what the audition button is for.

Sections that were kept because the transcript was unclear

When the transcriber could not hear a span cleanly, RoughCut keeps it and flags it. It is never silently deleted, and the row's reason pill is tinted pink rather than orange.

A banner above the list counts them: <n> sections couldn't be transcribed cleanly. They are kept in your cut — use ⏵ to listen and decide.

The two reason strings are Unclear transcription — listen before trusting and Transcription may have degraded here — listen. Both mean the same thing for your purposes: the words on screen may not match what you actually said, so judge that span by ear, not by reading it. If it is genuinely unusable, click Remove on that row.

Audition any range

Each row has a play/pause circle. It plays exactly the row's [start, end] range from your source media — audio only, with no video. The tooltips read Play this section's audio and Stop, and only one row plays at a time.

This is the fastest way to resolve a disagreement with the editor. A row whose transcript reads like a clean take may still be the one where your voice cracked.

Flip any decision

The button on each row is labelled with what it will do, not with the current state:

• Keep appears on a removed row. Clicking it reverses the proposed removal.
• Remove appears on a kept row. Clicking it excludes that segment — and it is also how you accept a Second Look suggestion.

Above the list, a counter reads Kept <n> · Removed <n> and updates live as you flip rows.

If you flip a row back to the decision the pipeline originally made, RoughCut discards the override rather than recording a no-op change. Undo every change and the job is exactly as it was.

The editor summary banner

A line near the top of the review tells you which editor produced these cuts:

• LLM editor: skipped — silence-only cut. — no AI pass ran, because Bad-take removal was off or no key was available. Every row you see came from silence detection.
• Editor: <Provider> (<model>) — <n> spans — the AI pass ran on that provider and model and proposed that many spans. The line may also carry the call's latency, and (cache hit) when the response was replayed from cache rather than re-requested.

If Second Look was on but could not complete, a separate orange banner says so and confirms that what you are looking at is the standard single-pass result.

Your changes are pending until you recompute

Flipping rows does not change the exported timeline on its own. The cut plan and the FCPXML are rewritten only when you click Recompute Cuts, which is enabled as soon as at least one decision differs from the pipeline's own.

Export before recomputing and you export the original cut. See Recompute after review changes.

Common mistakes

• Reading the transcript instead of listening. The transcript is a summary of a span, truncated at 220 characters, and on flagged rows it may be wrong. Audition anything you are unsure about.
• Assuming a hidden silence row is unreviewable. Switch to By category to see and reverse every silence and micro-island removal.
• Exporting after flipping rows without recomputing. The export reflects the last computed cut plan, not your pending toggles.
• Deleting flagged sections on sight. They are flagged because the transcript is unreliable, not because the footage is bad.
• Trusting a high percentage as proof. It is the editor's own confidence, not an accuracy guarantee.

Next steps

• Review cuts by category and confidence
• Recompute after review changes
• Export and import FCPXML