Adapting a public LiDAR barrow detector to a new region: specialist-start fine-tuning, an equal-update replay comparison, and evaluation choices that move apparent gains

Public deep-learning detectors for archaeological earthworks in airborne LiDAR are now available with released weights, and regional fine-tuning of such models has been reported before. This paper is a narrow case study of one released checkpoint, one class and one region. Under a specified decoder and a register-enriched sample of each district, fine-tuning the public ADAF barrow segmentation checkpoint on register barrows from three districts of Lower Saxony (Germany) increased registered-barrow recall relative to the unchanged checkpoint, under a total-proposal-budget benchmark with individual barrows as entities and one-to-one matching. Ten stochastic fits averaged +6.25 percentage points at 211 proposals (0.5 per km²) in Verden, the recipe-selection district, where the same runs had already been scored under an earlier metric (recall 0.169 for the unchanged checkpoint; all ten fits positive). In Oldenburg, a district that shares no training data but had been used in earlier work, including the development of the benchmark on saved proposals of earlier systems, and where this adapted model's output had not been evaluated before the new decision rule was frozen, the ten fits averaged +3.86 points at 225 proposals over a recall of 0.099, all ten positive, with a joint 95 % block-bootstrap lower bound of +1.37. With a common start, update count and recipe, a specified mixed-class replay package of nonlocal English, Welsh and Irish barrow and enclosure labels reduced Verden register recovery (−6.09 points against the unchanged checkpoint), whereas the locally selected barrow package improved it; the +12.34-point package difference combines improvement and deterioration and does not isolate geography, target class, exposure or recipe suitability. Rescoring fixed saved proposal sets under explicit entity, duplicate, budget and IGNORE conventions (IGNORE: register records that are neither targets nor background) changes measured model contrasts, including one earlier point-estimate sign. I identified a mismatch between the written IGNORE wording and its historical implementation, preserve the original results, and report three explicit readings as post-hoc sensitivity analyses: the adaptation contrast stays positive in every run, district, decoder and budget, while its magnitude and practical-threshold verdict vary (Oldenburg under the literal reading: +1.92 points, joint lower bound +0.62, below the +3-point practical criterion). The gain is attributable to the adaptation package as a whole (specialist start, local labels and recipe); neither the recipe alone nor geographical locality is isolated. The evidence concerns registered barrows in one region and previously used terrain; it is not evidence of archaeological discovery.