SwitchLogHow we measure
How we measure
SwitchLog is developed against the videos creators actually published, not against a taste panel. This page describes the method. It states no numbers: a figure is published only once it has been measured on more than one recording and the report says what it was measured on.
In plain words
Four steps, no formula needed.
Take one pair
A video you published and the raw recording it came from.
Line them up
Every second of the raw gets a label: you kept it, you trimmed it inside a take, you dropped it, or nobody can tell.
Let the engine cut blind
The engine cuts the same raw without ever seeing your edit.
Compare, second by second
Of what the engine kept, how much did you keep too? Of what you kept, how much did the engine find? What did it throw away that you had kept? And what would random picking have scored?
A worked example, with round numbers
- Say you published 20 minutes from a two-hour stream, and the engine also kept 20 minutes.
- 12 of the engine's 20 minutes are minutes you kept too. Precision: 60%.
- Those 12 minutes are 12 of your 20. Recall: 60%.
- Picking 20 minutes at random from 120 would hit your 20 about 17% of the time. That is the chance rate printed beside every precision.
- If the engine dropped 3 minutes of what you kept, that is reported as over-removal on its own, never folded into the score.
The pair
The unit of measurement is a pair: a creator's raw recording and the video they published from it. From the pair we build a truth track that labels every second of the raw as kept, trimmed, dropped, or unknown when nobody can vouch for it.
The comparison
The engine cuts the same raw footage without seeing the edit. Its choice is then compared with the truth, second by second and time-weighted: the share of chosen seconds the creator also kept, the share of the creator's seconds the engine found, and the overlap of the two as a whole.
What is reported separately
Removal is scored on its own, and the seconds the creator kept but the engine removed are reported as their own rate, never folded into a single score. Unknown seconds are excluded from every rate. Chance, the rate random choice would get, is printed beside every precision.
Personal profiles
A profile is fitted on a creator's pairs and checked on folds it did not see. When there is only one pair the report says so, because one recording is one sample.
What it does not prove
Agreement with an edit is not viewer satisfaction, and a shorter output is not a better one. Those are stated as limits in every report, and release notes must mark anything unmeasured as unmeasured.