Method
How it works, transparently.
NeuroVidz measures what is in a clip — the picture and the sound, second by second — and scores how well it hooks, holds and moves a viewer, with the edits to make. Nothing here is measured from a brain; every number comes from the clip, and each one shows its source.
Step 1
Measure every second
39 signals, one value per second, from the picture and the sound: cuts, motion (camera moves kept apart from moving subjects), faces and expressions, brightness and colour; speech, pace, voice pitch, music, beat and loudness.
- Loudness follows the broadcast standard (ITU-R BS.1770), measured as the clip will play after a platform evens loudness out.
- Speech is found by a neural voice detector, so music, tones and noise are not mistaken for talking.
- Speaking rate counts syllables inside speech only; beat strength is corrected for what noise produces by chance.
- Every measurement is tested on clips built so the right answer is known — see “How it is tested” below.
Step 2
Check what holds attention
The measurements are read against three parts, each a short checklist. Every check says what it measured, what good looks like, and the edit that would improve it.
- Hook — something happens at once; the picture changes early; the first word lands inside a second; a face early (when the clip has faces); the opening is as lively as the rest; it is loud enough.
- Hold — no dead air; the picture keeps changing; sound carries through; something new every few seconds; no part of the clip sags; it ends on something.
- Feel — a clear high point; intensity rises and falls; expressive delivery (voice, faces or music); more than one emotional beat.
Step 3
Read the story
A language model reads what is said and shown and answers seven plain questions — promise, payoff, clarity, specifics, surprise, stakes, ending — each with the moment it is judging. It runs at a fixed setting, a “yes” needs evidence, and the answers are counted by us, not by the model.
Step 4
Score it, and list the fixes
Each part is the average of its checks, and the score is the average of the parts — the three measured ones and Story. A check that cannot apply to a clip (no faces, no speech, expression reading switched off) is left out and named, never counted as a failure. The fixes are ranked in the order a viewer meets the problems, each with the moment and the change it would make to the score.
How it is tested
Known answers, and the same score twice
- A metronome at 120 BPM reads 120 BPM; major and minor chords read major and minor; a test tone, white noise and chords read 0% speech; a camera pan reads as a camera move.
- The same voice 6 dB quieter reads the same pace, pitch and speech share; loudness itself moves by the 6 dB.
- On 24 real clips, the same file scored twice gives the same score; 2 seconds of dead air added to the start lowers the Hook part on every one.
- These checks are part of the test suite for the measurement code.
What it is not
- Not a brain scan. The 3D brain in a result is a model of the measured signals — a way to see them, not a recording of anyone's brain.
- Not a view-count predictor. The checks follow what research and platforms say holds attention; the four parts are weighted equally until outcome data can justify other weights.
- Not a judge of taste. It reads structure — timing, change, delivery, story beats — not whether a joke is funny to you.
