Whole-Path Hit Rates and Error in Sparse Range–Bearing Forecasting: research data, code and manuscript

CORRECTION (1 October 2026): Files in this version are restricted because its manuscript PDFs and source included a journal-specific template unsuitable for public posting. The corrected open version, with neutral manuscript formatting and identical scientific results and evidence archive bytes, is available at https://doi.org/10.5281/zenodo.23074661 (v17-neutral). This record and DOI are retained for provenance. Please use the corrected open version. Research archive accompanying Whole-Path Hit Rates and Error in Sparse Range–Bearing Forecasting (v17-final), by Xu Zhang and Xianjian Zhang. The package includes the manuscript and supplement, multi-file LaTeX source, synthetic observations, model checkpoints, original evaluation records, observation-only replay, and separately labelled post-review extensions. Finite trajectory sets can favor different error and hit-rate objectives under sparse range–bearing sensing. We evaluate three learned model seeds on 6000 causal synthetic histories. Direct set-retention training reduces Top4 selection failure at 60 m by 1.66 percentage points relative to error-based ranking, but its mean-error effect depends on budget. Fixed-slot controls and a no-ID ablation limit interpretation as better per-history ranking. Analytic banks remain weaker than the learned eight-path bank. This is a bounded synthetic comparison, not real-platform validation or a deployment recommendation. Reproduction: original primary analyses and post-review extensions remain distinct. SHA-256 file manifests are included. Prediction replay was verified in a relocated directory using the recorded dependencies (Python 3.12, NumPy 2.5.3, PyTorch 2.14.0); fresh-install replication and independent full retraining have not been validated. Original training/selection role caches are not all duplicated and must be regenerated for full retraining. Author-only correspondence and review reports are excluded. Licensing: original code is MIT; original data, checkpoints, manuscript, supplementary material and archive documentation are CC BY 4.0. These licenses apply to different components, not as alternative licenses for every file. Third-party materials, including the MDPI template, retain their existing licenses. See LICENSES.txt and LICENSE-MIT.txt inside the archive. This is a research archive, not an accepted journal publication. Download note: the v15 numerical-evidence ZIP and v16 review-extension ZIP are provided as 6 and 11 binary parts (at most 20 MiB each). Download all parts plus SPLIT_ARCHIVES.json and reassemble_archives.py to one directory and run python3 reassemble_archives.py. The standard-library-only script verifies each part and reconstructs the original ZIPs with SHA-256 checks. Other ZIPs are directly usable. See README_ARCHIVE.md for details.