New Brunswick CORE4 structural core late-successional forest: 10 m probability surface, bootstrap uncertainty and extrapolation flag, trained on the licensed provincial inventory (2024 AlphaEarth embeddings)

A 10 m wall-to-wall probability surface for the CORE4 structural core late-successional condition across New Brunswick forestland, with a matching block-bootstrap standard deviation layer and a training-envelope extrapolation flag. Predictions come from a random forest over the 64 Google AlphaEarth Foundations 2024 satellite embedding bands and a Hansen v1.12 time-since-disturbance covariate, fitted on the 9,042 plot Configuration D basis of the provincial Continuous Land Inventory as delivered under licence, scored with the CORE4 instrument at three of four comparably measured axes (large tree basal area, maturity on maximum diameter, structural diversity, canopy). The mosaicked forestland share is 10.65 percent (bootstrap percentile interval 10.06 to 11.82 across 20 forests), which falls inside the independent design-based estimate of 9.37 percent (95 percent interval 7.71 to 11.03) on the same licensed frame, with no calibration step. The surface passed an eleven-gate acceptance suite spanning location, value range, total magnitude and cross-layer consistency, a plot-level assembly check against a permutation null, and a separate post-acceptance stress test. Why this replaces the earlier New Brunswick surface. A 16 September 2026 surface trained on labels scored from the public CLI release read about half the licensed rate, because the public tree extract misses about half of the plots the licensed records score as structural core on the same plots. Refitting on the licensed labels with nothing else changed is this deposit. The two surfaces correlate at r = 0.80 over 667 million cells and this one sits above the earlier one on 95 percent of them. Intended use. This is a continuous probability surface, not a classified map. Out-of-fold block calibration is attenuated (predicted-on-observed slope 0.262, interval 0.062 to 0.438), which compresses block means toward the middle of the range, leaves the provincial total close to correct and biases a fixed threshold. Users thresholding this surface should calibrate the cut against a design-based estimate for their own domain rather than adopting a nominal value. Ensemble SD rises with probability, so the uncertainty layer is the instrument for reading the high end. The fitted random forest and its 20-model block-bootstrap ensemble are included so the surface can be regenerated and audited; a memorisation measurement on the forest is deposited with them. No plot-level inventory records, plot coordinates, per-plot embeddings, per-holder or per-parcel outputs are included in this deposit. Release of this aggregate product from the licensed inventory was permitted by the New Brunswick partners in September 2026.