What the engine does, and what it does not claim.

1. The video becomes passages

SlateRoom transcribes the upload with word timings, finds shot changes and splits the video into passages. Every later step works on those passages, so every change can point at exact words and seconds.

2. The audience comes from real accounts

An audience is built from accounts you choose: your own channel, channels whose viewers are like yours, or channels you want to reach. SlateRoom reads their public comments and, where it has access, measured demographics, and writes a profile. Each claim in the profile is marked as observed or inferred, and the profile says how many comments and videos it rests on.

3. The read: changes tied to the source

The engine compares what the video promises with what it delivers for that audience, and proposes changes: cut, move, keep, or record something new. Each change quotes the passage it is about with its source times. The model cannot invent a timecode: times come from the transcript, not from the model.

4. A second read: what people would talk about

A simulated community, built from the audience's own public comments, reacts to the transcript. It shows which moments draw talk and what people would argue about. It is labeled simulated everywhere it appears.

In our tests on held-out episodes from two shows, a community built from a show's own comments matched the topics of the real comment section more closely than a community with no history. Across three episodes with measured viewing, how much the community talked about a moment did not predict how many viewers stayed. So SlateRoom never uses simulated talk to forecast retention or to justify a change.

5. Predicted and measured, kept apart

Before publishing, the expected viewing line is an estimate made from how viewers left the channel's other videos of a similar length. After publishing, the actual line comes from YouTube Analytics where SlateRoom has access to the channel. The two are drawn separately and labeled. When results arrive, the read shows whether the drops it flagged are where viewers actually left.

What we can defend

• Every change points at real words and seconds in your video.
• Every audience claim says what it rests on.
• Simulated talk, estimates and measurements are labeled and never mixed.
• A read that finds nothing worth changing says so.

What we do not claim yet

• That following a read raises views, watch time or click-through. We have not run that test.
• That thumbnail recommendations beat other options with real viewers. The pick is an engine ranking; use YouTube's Test & Compare to check it.
• That comments represent the whole audience. People who comment are a small and sometimes different group.
• That exports import cleanly into every editing app. Imports are still being verified.
• That the engine understands every community equally well. It has been used most on interview, podcast and relationship content.