The cluster is a hypothesis, not a finding
Strip the top line off the leaderboard and a quieter pattern remains. Six frontier models, built by four competing labs, were reported between 76.2 and 80.6 on the same task, a spread of 4.4 points by our arithmetic. That is the company's reported evaluation. Whether the scores are comparable to each other, and whether a spread that size is significant, is unverified.
Rehearsals says the models assume people are less aware, more frugal and lazier than the survey data shows. Its chosen illustrations: models expect most over-65s not to watch short-form video, while 80% of the surveyed group had watched in the last week. Models put the share of 55-to-64-year-olds who would keep a plan after a $4 price rise at half; the survey says 70%. Models put parents of toddlers doing the cooking at 30%; the survey says about 66%. Models put older Midwest shoppers paying by phone tap at 5%; the survey says 25%.
These are the release's selected illustrations, picked to make its point, and they are not independently checked here. Within them, the errors run one way: underestimating technology adoption and overestimating price sensitivity. Read that as a hypothesis worth testing against real survey data, not as an established finding about frontier models. It does not become a finding until someone tests it on material Rehearsals did not select.
A home game, described in the release itself
The 94.1 needs more handling, because the release is unusually candid about how it was produced.
Rehearsals built this benchmark as the scorecard for its own product. The product is a system for simulating survey panels: Rehearsals sells predictions of marketing outcomes, made by AI twins of real people, in 30 minutes. The benchmark tests exactly that skill, predicting the distribution of answers a human panel gave. Engram was trained on hundreds of thousands of hours of interviews gathered by the same research platform that gathered the benchmark's interviews. The release is careful to say Engram was not trained on benchmark content and learned from a separate set of human data.
That separation, if accurate, is relevant. It is not the same as demonstrated contamination, and this article does not claim leakage. What it does establish is a potential advantage: a model raised on a platform's interview data starts a test built from that platform's interview data knowing the source. The populations, methods and distributions behind the two sets cannot be assessed from the release, so the size of any advantage cannot be assessed either.
Then there is verification. 80% of the benchmark is a private holdout, and the disclaimer confirms most of the data is private, which is a legitimate defence against training contamination and also means outsiders cannot reproduce the full evaluation from released data alone. Engram itself is internal, previewed with select partners, so nobody outside Rehearsals and its partners can repeat the headline comparison at all. And the document carrying all of this is a paid press release, a format in which the subject writes the text and pays for distribution. That is relevant context. It is not proof the score is false. It makes the score uncheckable from outside, which is a different problem, and the one that matters if the number starts travelling.
What the Kaggle hosting does and does not establish
Rehearsals says the bench is hosted on Kaggle so any company or individual can score their latest model, and that the results were conducted by Rehearsals and Google DeepMind using the Kaggle platform. Its own disclaimer adds that naming DeepMind and Kaggle does not constitute endorsement, sponsorship or affiliation. Checks for this article found no primary DeepMind statement confirming a role, and the Kaggle and benchmark pages checked did not resolve what an outsider can actually submit, whether evaluation code is visible, or exactly what portion of the data is public. The release alone does not establish those points.
Even taking the hosting claim at face value, submitting a model for a score is not independently reproducing the evaluation. With 80% of the data private, an outside entrant can be scored without being able to re-run the full evaluation, audit the private holdout, or repeat the comparison against Engram, which is not publicly runnable. The extent of external verification is unclear from the release.
So the practical reading is two-tier. The frontier spread, 76.2 to 80.6 as reported, and the one-way errors in the release's examples are hypotheses from the company's own selected material. The 13.5-point gap is a claim awaiting an outside run the release gives no way to perform. Rehearsals says it expects a broader version, with more participants and four more countries, before the end of 2026. That is a forecast, not a confirmed release date, and a second version would be a better test only if more of the data behind it can be inspected.
Sources
- [1] Rehearsals (Deep Prediction Inc.), "Rehearsals Launches Human Behavior Bench With Google DeepMind and Kaggle to Measure How Well AI Models Understand Humanity", EZ Newswire via FinanceWire, 20 October 2026 — primary source for the launch, the scores, the dataset, Engram's training description, the partnership framing and the disclaimers; the release states it is a paid press releaseRead source
- [2] Syndicated copy of the same release, RushPR News, 20 October 2026 — identical text; used to confirm the release wording, not as independent corroborationRead source