All stem samples were collected before dawn. For each tree, a large branch (\($\approx$\)0.8 to \($\approx$\)2 m length; \($\approx$\)10 to \($\approx$\)25 mm base diameter) was removed from the canopy using a telescopic pruner. The cut edge of each branch was immediately sealed with parafilm and electric tape, and sealed branches were transported to our base (an air-conditioned, darkened room) within 30 minutes of cutting. Upon arrival, leaves were cut with a sharp razor blade for pre-dawn leaf water potential (LWP) measurements (see description below). Two separate stems from the same branch were then cut, sealed at both ends with parafilm, wrapped in cling wrap, and double Ziploc bagged. Stems were subsequently stored at 4°C for later water extraction through CVD and Cavitron centrifugation methods. The sampled stems typically ranged from 210 to 260 mm in length and 6 to 13 mm in diameter.

We used a Model-1000 pressure chamber (PMS Instrument Company, United States) with nitrogen gas and a mounted eye lens for the LWP measurements. Two replicate measurements were made for each branch sample, plus a third measurement when the difference between the first two measurements was > 10%. All LWP measurements were finalised within two hours of sampling. For each sample we report the average of the two to three measurements.

To characterise the spectrum of isotopic compositions of potential source waters, we sampled both groundwater and soil water at various locations and depths within and around Elsey National Park. We obtained groundwater from five observation bores and soil water from six soil cores. For groundwater, we used a submersible pump (Tornado, Proactive, USA) except at one site where we used a pre-installed solar-powered pump, and collected samples once three bore volumes had been purged and/or once pH and conductivity measurements had stabilised. For soil samples, we used a hand auger at five sites to extract shallow soil (maximum depth of 2.0 m), while at one site, we used a small drill rig to extract deeper soil horizons (maximum depth of 5.4 m). We collected a total of 36 soil samples at depths ranging from 0.1 to 5.4 m. Each sample comprised approximately 100 to 300g of soil material, was sealed in double Ziplock bags with minimised headspace, and kept at 4°C until further analysis. Additionally, a local meteoric water line was obtained from 18 rainfall samples collected on site between 2019 and 2021 (Lamontagne et al., in prep.).

To extract xylem water from stem samples, we used a standard 270-mm diameter Cavitron manufactured by DG-Meca (France) and fitted to an ultracentrifuge (Avanti J-E, Beckman Coulter, United States) at Charles Darwin University. We broadly followed the method outlined in Barbeta et al. (2022). Briefly, small plastic containers were inserted into each end of the stem and sealed with parafilm to collect xylem water. Samples were spun for two minutes at speeds ranging from 3,000 to 9,000 rpm, corresponding to xylem pressures of –0.57 to –6.40 MPa based on stem length (Alder et al., 1997; Cochard, 2002). Extracted water was then collected from the containers using a micropipette, filtered through 0.45 μm and stored in 2 mL glass vials fitted with 0.1 mL micro-inserts. Samples were then preserved at 4°C until analysis. Stem samples were weighed before and after centrifugation, oven-dried at 105°C for 24h after centrifugation and reweighed to determine the relative stem water content (RSWC). A more detailed description of our operating procedure is provided in the Supplementary Information.

Bulk stem water and soil water were extracted at the West Australian Biogeochemistry Centre (WABC), University of Western Australia, following the CVD procedure outlined in West et al. (2006). Samples were fully frozen using liquid nitrogen, after which they were subjected to a vacuum with pressure < 10 Pa. Frozen samples were then heated under vacuum conditions, causing water vapour to be collected in a liquid nitrogen cold trap. Extraction times were set at 60 and 90 min for soil and stem samples, respectively, aligning with the recommendations of West et al. (2006). To ensure the quality and accuracy of the extraction process, water was also extracted from four different standards using the same procedure.

All extracted water samples were analysed for oxygen and hydrogen isotopic ratios (δ¹⁸O and δD) at the WABC using a cavity ring-down spectrometer (Picarro Inc., model L2130-I) fitted with a micro-combustion module to remove organic compounds that may be present in extracted water. The raw isotopic values are expressed relative to VSMOW and are reported in per mil (\($\textperthousand$\)). According to analyses on replicate stem samples, overall precision for the CVD extraction and measurement procedure was ±0.5\($\textperthousand$\) and ±3.0\($\textperthousand$\) for δ¹⁸O and δD, respectively. Cavitron-extracted xylem water and groundwater samples had a precision of ±0.1\($\textperthousand$\) and ±0.5\($\textperthousand$\) for δ¹⁸O and δD, respectively.

Statistical analyses and plotting were conducted using MATLAB R2022a. We define deuterium bias as the difference between δD of CVD-derived bulk stem water and that of Cavitron-derived xylem water of the same branch:

δD_bias = δD_bulk – δD_xylem (1)

We define deuterium offset as the difference between δD of xylem or bulk stem water and their expected δD based on the source water line:

δD_offset = δD_{xylem or bulk} – δD_source (2)

with δD_source = α_swl * δ¹⁸O_{xylem or bulk} + β_swl (3)

where α_swl and β_swl are the slope and intercept of the source water line, respectively.

To test whether the means of the δD offsets of xylem and bulk stem water were significantly different from zero, we used a one-sample t-test (MATLAB function ttest). To test whether the difference between the δD offsets of xylem and bulk stem water was significant, we used the non-parametric two-sample Kolmogorov-Smirnov test (MATLAB function kstest2). Unlike other tests (e.g. paired t-test) that only compare the means of each group, this test evaluates the entire distribution of each group. To test cross-species differences in mean δD bias, we used a Kruskal-Wallis test (MATLAB function kruskalwallis) followed by a Dunn-Sidák post-hoc test for pairwise comparison (MATLAB function multcompare).

To test the potential effect of tree species, RSWC, pre-dawn LWP and xylem water isotopic composition on the CVD-induced δD bias, we used linear mixed-effects models (MATLAB function fitlme). We used species, pre-dawn LWP, δD of xylem water and RSWC as predictor variables and δD bias as the response variable. In addition, we included sampling site as a random effect in the mixed-effects model.