PERSEUS libcbm vs GCBM cross-state intercomparison (CONUS)

Scientific finding. Two implementations of the CBM-CFS3 forest carbon model family, libcbm (Canadian Forest Service Python/C++ reimplementation) and GCBM (the moja FLINT spatially explicit engine), are compared across six US states under matched parameter conventions. The year-five carbon density gap is approximately +24%, with libcbm/GCBM ratios between 0.74 and 0.80 in Washington, Minnesota, Indiana, Maine, and Oregon, and a single warm-donor outlier in Georgia at 1.05. Replacing the uniform forest-type stratification used in legacy parity runs with a stratification anchored on FIA EXPNS expansion factors (the canonical FIA Total Area Estimator) adds +6,760 TgC to the CONUS year-five carbon stock at n=39 states, a +14% shift on 246.5 Mha. We conclude that inventory stratification, not engine implementation, dominates cross-model uncertainty for this generation of CBM-CFS3 carbon models. Regional DOM-pool fingerprint. The deposit further documents a regional dead organic matter pattern that distinguishes cool-moist climates: mean libcbm slow soil carbon under B1.3 FIA EXPNS is 184 Mg/ha in the Pacific Northwest, 143 in the Atlantic Maritime Northeast, and 105 to 119 Mg/ha across the South, Lake States, and Mountain West. A Q10 mean annual temperature sweep on Oregon (MAT 4 to 13 C) shows the slow-soil overshoot is established during spinup rather than emerging from runtime Q10 scaling: libcbm/GCBM stays at 1.35 even at MAT 13 C. Spinup climate is therefore a candidate for recalibration in any future EPA GHG inventory or NCASI national-scale benchmark that adopts libcbm as a Tier 3 implementation. Contents. 48-state libcbm year-five pool outputs under canonical B1.3 FIA EXPNS inventory; 6-state libcbm vs GCBM density gap matrices under both B1.1 v6 parity and B1.3 FIA EXPNS; per-pool and per-DOM-pool decompositions; F3 Q10 MAT sensitivity sweeps for Georgia and Oregon; and the five publication figures supporting the methods finding. Pipeline. All libcbm outputs were produced with the GCBM2hpc pipeline (github.com/holoros/GCBM2hpc) on the Ohio Supercomputer Center Cardinal cluster (allocation PUOM0008). Six-state GCBM aggregates were produced spatially via moja FLINT containerized GCBM at 1-degree WGS84 tiles. Auxiliary CONUS layers: TreeMap 2022, LCMS v2024, FIA DataMart, NOAA 1991-2020 climate normals, EPA Level III ecoregions. Companion methods note: holoros.github.io/perseus-forest-intelligence/methods/inventory-stratification/