If you came in doubtful, good. These are the questions a careful reader asks, answered without dodging. Where the honest answer is "we don't know yet, we're testing it," that is what you'll read.
Soma predicts the cortical activation a video would produce in an average viewer, straight from the file, using Meta's public TRIBE v2 model. Off that predicted activation we read an attention arc and a coarse affect arc (valence and arousal).
The prediction of activation is the validated part. The attention read is validating now. The affect read is a labeled hypothesis. The demo shows all three with those exact badges.
No. There are no people and no hardware in the loop. Soma runs a model that was trained on real fMRI and predicts the response an average viewer would have. It is a prediction from the file, not a measurement of your audience.
Short-form video: TikTok, Reels, Shorts, and paid-social ad cuts. It reads from the file in minutes, so you can run every variant instead of only the one you could afford to put through a panel. Pricing is being set for performance and creative teams; it will sit far below a human panel study.
The encoder is validated by Meta against real scans. The attention read-out is being tested right now against public human data (TVSum), and we report the result either way, including a null. The affect read is a labeled proxy, not a decoder. We would rather show you a real null than a pretty number that isn't earned. The science page lays out the full method.
Other brain-AI tools quote a “92%” figure and attribute it to Meta. We can't find it in any Meta publication, so we don't use it. Meta's TRIBE actually reports a mean correlation around 0.21 across ~1,000 cortical regions on held-out data (about half the measurable ceiling). Either way it only measures video → brain activation versus real fMRI — not whether activation predicts whether someone keeps watching.
That downstream step is a separate question, and it is the one we are testing in public.
It cannot name a specific emotion from activation. A spike could be interest, confusion, mild alarm, or noise, and only behavior settles which. It does not yet predict calibrated retention; that is a roadmap rung earned by a held-out test, not a shipped feature.
We also disclose a known negative result: whole-brain activation does not predict which parts of a YouTube video get replayed, so the whole-cortex version of our signal is the baseline to beat, not the win.
An LLM role-playing a viewer imagines what a person might say, and it leans toward agreeable, plausible answers. There is no biology under it. Soma predicts the neural response from a model trained on real brains, and the same input gives the same output run to run. Different signal, grounded in a different place. See the comparison for the full field.
They could download the same public model tomorrow. The reason they are slow to is that a transparent, from-the-file prediction undercuts their own pitch: recruit a panel, trust our score, pay per study. It is a business-model conflict, not an engineering gap. Incumbents rarely ship the thing that commoditizes their pricing.
Correct, and we say so plainly: the encoder is not a moat for us or anyone. The moat is three things the model can't give you. First, the honest validation almost no one bothers to do. Second, the product and the daily workflow. Third, a data flywheel of real ads paired with real reactions and retention from our partners, which a public-model competitor can never scrape. Whoever earns real validation and gathers real outcome data first wins the honest version of this race.
Yes. The encoder is Meta's public TRIBE v2, so the first step is reproducible by anyone. Our test plans are written down before we look at results, nulls are reported like any other outcome, and this site ships with no build step, so the code behind the honesty labels is readable in your browser. When a real validation number lands, it is filled in from the data, not typed by hand.