What we measure, what we know, and how we're testing it

PrePost runs Meta FAIR's TRIBE v2, a model trained on brain scans (fMRI) of people watching and listening to media. Given a new video, it predicts the brain response an average viewer would likely have, once per second, across the cortex. We summarize that into seven brain areas and a timeline. It is a diagnostic: it shows which moments the model expects to land and which to fall flat. It is not yet a forecast of views or retention, and we don't claim it is.

• A per-second timelineof the predicted whole-cortex response. The units are the model's own z-scored scale, so compare moments within a clip, not raw numbers across clips.
• Seven brain areas, so you can see whether a moment works through the visuals, the voice, the people on screen or the story.
• Plain-English notes with timestamps, computed from those numbers: the weakest and strongest stretches, how each shot compares, which area does the most work.

Areas are Destrieux-atlas approximations on the fsaverage surface. “Roughly tracks” is a simplification: no brain area does one job.

• The model itself is well validated at what it was built for: predicting brain responses to movies and audio. It placed first in the Algonauts 2025 challenge.
• Our first pilot did not show a reliable link to engagement. Across 24 videos, the correlation between the read and engagement was r = 0.25 with p = 0.24. That is weak, and with only 24 videos it could easily be chance. It neither proves nor rules out a real effect.
• An outside study points the same way.Sahu & Pandey (arXiv:2607.01400, 2026) found TRIBE's predicted signal did not predict YouTube re-watching.

Put together: a brain response the model predicts well is not the same thing as a video that performs well, and so far nobody has shown the second follows from the first for short-form video. That is the gap our test is built to close, or to confirm can't be closed.

The question: does adding the predicted brain response to a content-only model (visuals, loudness, motion, transcript, length) improve prediction of how many people watch past the first three seconds? If the brain read only restates things a cheaper model already sees, it adds nothing, and the test is designed to catch that.

• Stimuli: 18 short base videos shot by our team, each cut with 4 different openings, for 72 clips. Only the opening changes inside each set.
• Delivery: each clip runs as its own ad with an equal daily budget, all starting and stopping together for at least three days, targeting about 1,000 impressions per clip (about $580 in ad spend).
• Outcome: hook rate, meaning 3-second plays divided by impressions, as the platform reports it.
• Blind scoring: every clip is scored before any performance data exists.
• Bars set in advance: the brain read has to track hook rate within each set of openings, and the combined model has to beat the content-only model by a fixed margin under cross-validation and a permutation test. A negative control, the same pipeline with random features in place of the brain read, must fail. If it passes, the result is void.
• Power: simulated before freezing, the design detects a moderate effect (rho 0.4) with about 90% power.

The four possible verdicts, decided before the data

Whatever the verdict, we publish it, with the effect size, its confidence interval and the population it applies to.

• It predicts an average brain from the scan data the model learned on, not your audience.
• Nobody is scanned. These are model predictions, not measurements.
• TRIBE v2 is licensed CC BY-NC 4.0, so PrePost is non-commercial research for now. See the license notice.