Method

Method and limits

By the neuro-viral team · Updated 12 Oct 2026

What the model does, how the chart is built, what we tested, and where the product stops. If a number on this site is not explained here, tell us at hello@vincli.com.

What the model is

neuro-viral uses a brain-encoding model: a model trained on fMRI recordings of people watching and listening to video. It learns the relationship between what is on screen and how the cortex responded. For a new video it reads the frames, the soundtrack and the spoken words on one timeline and predicts the average response.

What comes out

• For every second, a predicted level of activity at 20,484 points across the cortical surface.
• Those points are grouped into 360 labelled regions (the Glasser HCP-MMP1 atlas) for the 3D map.
• The same prediction is summarised as one second-by-second chart.

No viewer is measured at any step. The prediction comes from the video alone.

How to read the chart

• Each chart uses its own scale. The line is drawn against the range of that one video. A level in one video cannot be compared with a level in another.
• The strongest second leaves out the first three. Every video starts with an onset response, so the first three seconds are not counted when the strongest second is named.
• A fall is a change, not a verdict. A quiet stretch can be a deliberate choice. The chart shows where to look, and the editor decides.

The trust score

Each second carries a trust score from 0 to 100. It combines three things: how accurately the model predicts the brain network that leads that second, measured on recordings it was not trained on; how far your video is from the kind of material the model was trained on; and how much separate copies of the model agree. When one of the three could not be measured for a scan, the report says so. A low score means that second should be read with caution.

What we tested, and what failed

We tested whether a predicted brain response helps forecast how an audience responds to a video. It did not.

That test used a brain-encoding model of the same kind, in June 2026. An independent study reached the same conclusion: Sahu and Pandey (2026) compared a predicted brain signal with the "most replayed" data of 48 YouTube videos and found a partial correlation of 0.058, with an interval that includes zero (arXiv 2607.01400).

So the product does not claim to forecast views, watch time, clicks or sales.

What it is good for

The prediction is specific about time. It shows which seconds of one video stand above or below the rest of that video, and how a revised cut differs from the original, second by second. That is a tool for an edit review. It does not replace a live test.

What is not modelled

• Regions below the cortex. The areas most linked to reward, emotion and memory sit deeper in the brain and are not part of the prediction.
• Your audience. The prediction is for an average viewer, not your customers or subscribers.
• Everything outside the video. Targeting, timing, offer and distribution are invisible to it.
• Unfamiliar material. The model is more reliable on video that resembles what it was trained on. The trust score reflects this.