Cross-border recovery of registered burial mounds by an adapted public LiDAR detector at fixed review workloads: a sealed Dutch test, a confirmed Danish system, and seed-level recipe repeatability

Public deep-learning detectors for archaeological earthworks in airborne laser scanning (ALS) data now come with released weights, and adapting them to new regions has been reported. Whether such adaptations hold up across national borders, under evaluation fixed in advance, has received less attention. We report three linked results for round barrows. (1) Dutch transfer of one checkpoint. A checkpoint adapted from the public ADAF barrow model on 852 Lower Saxony register barrows was applied with no Dutch training data. In sealed Dutch test provinces it recovered 623 of 932 registered barrows. The public checkpoint recovered 516 at a matched workload: within each province both models kept the same number of top-ranked proposals, equal to the public model's whole output there (3,718 in total, about 2.3 per km² overall). The difference was +11.5 percentage points, Bonferroni-simultaneous 95 % interval [+7.4, +15.5]. This reproduced the +11.8 points measured earlier, under the same rule, on separate Dutch evaluation provinces. (2) Danish comparison of two fixed systems. The test used 171 barrow-dense West Jutland cores (4,170 km²), holding 5,273 protected round barrows, which no project model had processed before this comparison. Both systems used proximity re-ranking under a common cap of 2 proposals per km². A two-checkpoint system trained with a revised recipe recovered 84.2 % of barrows, against 78.9 % for the adapted checkpoint: +5.33 points, 95 % core-clustered interval [+4.39, +6.27]. (3) Danish comparison of two training recipes. Afterwards, on the same, already-scored cores and cap, single models from the revised recipe exceeded single models from the earlier recipe by +4.47 points, averaged over nine matched training-seed pairs (95 % interval including a seed-variance floor [+3.07, +5.88]). The difference was positive in all nine pairs. This tests training repeatability on one geography, not a further geographic confirmation. The Dutch test and both Danish comparisons were specified and committed before any of their evaluation terrain was read, and each frozen primary analysis was scored once. Independent computational recounts reproduced the numbers. The two size patterns use different measures. By register footprint, the adapted checkpoint's advantage over the public model was concentrated in small and mid-size Dutch mounds (descriptive, at unequal workloads); by local relief (a post-hoc diagnostic), the Danish system's advantage over the adapted checkpoint rose with relief, and no improvement was demonstrated below about 0.5 m. At matched workload, the project's English reference model recovered 15.1 points fewer Dutch barrows than the public barrow specialist. All results measure recovery of registered monuments at workloads defined in proposal counts. They are not estimates of discovery precision and not field validation.