On September 17, 2026, Figure AI announced that its Helix 2.5 neural network entered 30 Bay Area homes it had never encountered, attempted 420 household tasks across three behaviors, and completed 56% successfully without any data collected in those homes, while a comparable system trained on the same task data without Figure's pretraining strategy managed 9%. The sixfold gap stemmed from one variable: whether the model began from a checkpoint pretrained on Index, Figure's global-scale human behavior dataset, or from random weights, using two policies with identical architectures and training procedures. Figure also claimed the improvement is predictable: training four versions on increasingly large portions of Index while doubling the data each time produced a smooth decline in held-out action-prediction loss, and the team used only the smaller runs to forecast the largest run's loss to four decimal places before it began. In the test, Figure rented 30 homes across the Bay Area, held evaluation objects (toys, towels, and bedding) aside, and confirmed through AI review and human inspection that none appeared in Helix 2.5's task-specification data. A single fixed model checkpoint ran across all 30 homes, with no weights tuned on arrival.