How we validate drop-off detection

What we tested

When Retensis Vision 2 analyses a video, it flags specific moments on the timeline. This test answers one question about those moments: are they anchored to the structure of your edit, or could the same timestamps have come from anywhere?

A tool that scatters plausible looking timestamps across a video is indistinguishable from one that reads it, unless somebody measures the difference. So we measured it.

How we tested it

For every flagged moment we measured the distance to the nearest cut in the edit. A tool reading the video should land close to cuts more often than a tool guessing would. On its own that number means nothing, so we compared it against two baselines that stand in for chance:

• The first keeps the video and its moments exactly as they are, and rearranges where those moments sit inside it. It preserves how many moments there are and how the video is cut, so anything it beats is a real effect rather than an artefact of either.
• The second takes the moments from a different video of comparable length. It answers the harder question: are these moments about this video specifically, or would any set of moments have scored the same?

Results

A lower distance is a better result. The flagged moments sit closer to real cuts than either baseline, and the odds columns give the probability that chance alone would produce a result this good.

What this means

The moments Retensis flags are tied to the structure of your edit. They are not distributed at random, and they are not generic timestamps that would fit any video of that length, because both of those are exactly what the two baselines represent and both were beaten.

The practical consequence is that a flagged moment is worth opening. It points at a real decision in the edit, so the timestamp is a place to look rather than a number to take on faith.

Scope

This test measures one thing: how closely the moments Retensis flags align with the cuts in your edit. Every figure above is a distance, in seconds, from a flagged moment to the nearest cut.

The run published here was produced by Retensis Vision 2 and executed on 2026-09-19. Results are recorded per engine generation, so a future generation is measured on its own terms rather than inheriting these numbers.