BRANCHSNV — Validation and reproducibility archive v0.3.1 (including Experiment 07)

This is version 0.3.1 of the BRANCHSNV validation and reproducibility dataset. It retains the original v0.3.0 benchmark evidence and adds the post-hoc analyses supporting Supplementary Table S8, together with their frozen dependencies and reproducibility protocol. The release preserves the three original v0.2.0 validation files, documenting the validation analysis completed on 06 October 2026 and a separately executed reproduction completed on 21 October 2026. Dataset version 0.3.1 is distinct from the associated BRANCHSNV software release v0.2.0. Experiment 07 - Comparative benchmark Experiment 07 comprises 150 simulated phylogenetic scenarios spanning five mutation regimes, three phylogeny sizes (32, 128, and 512 tips), complete and masked observation conditions, and five replicates. The dataset includes 1,350 method–scenario outputs from BRANCHSNV and eight external workflows: ARPIP, FastML, HomoplasyFinder, PAML, PastML, POUTINE, SNPPar, and TreeTime. The original seven gzip-compressed TAR archives preserve canonical comparator observations, separately sealed simulation truth, BRANCHSNV outputs, the original 300 FastML/TreeTime executions, 900 additional external comparator executions, independent scoring evidence, and nine selected prospective same-host native replays. These original archives contain 30,019 SHA-256-inventoried regular files, approximately 2.32 GB compressed. A further 152 symlinks to external POUTINE compiled-runtime locations were deliberately excluded and documented. All nine selected native replay pairs generated byte-identical prediction files, reproducing 226 branch-event and 58 recurrent-site predictions. This provides targeted same-host reproducibility evidence, not independent clean-environment reproduction of all comparator executions. Supplementary Table S8 - Post-hoc analyses (added in v0.3.1) An additional archive provides two analyses of BRANCHSNV's branch-event predictions: - S8a - Error attribution: Case-level classification of strict true positives, false negatives, false positives, and ambiguous predictions across all 150 scenarios and 9,600 simulated truth events.- S8b - Event identifiability: Truth-replay and unit-cost parsimony diagnostics across 60 convergence and reversal scenarios, comprising 4,200 truth events at 2,400 sites. The supplementary archive contains the original analysis scripts, results, frozen scoring and validation dependencies, SHA-256 manifests, and an executable reproduction protocol. Reproducibility: The S8 analyses were independently re-executed from an extracted supplementary archive using canonical inputs retrieved from Zenodo and exact public Git revisions. All 1,800 required canonical files were reconstructed and verified against their published SHA-256 identities. Both analyses regenerated their original outputs byte-for-byte. The S8 analyses are explanatory post-hoc diagnostics. They do not modify the original predictions, comparator executions, primary scoring, or benchmark conclusions. Interpretation and limitations Simulator truth remained inaccessible during inference and was used subsequently for authorised scoring and post-hoc diagnostics. The benchmark evaluates conditional latent-history recovery using phylogenetic branch lengths derived from realised simulated events. Branch-event placement and recurrent-site identification are distinct scoring endpoints. Ambiguous prediction compatibility is not counted as a strict true positive. The S8 identifiability analysis evaluates compatibility under a specified unit-cost minimum-change model and does not establish biological impossibility. External software dependencies, event identifiability, and undefined zero-denominator F1 scores remain important interpretative limitations. Detailed file manifests, archive checksums, verification attestations, provenance statements, and reproduction protocols accompany the dataset. Source repositories Frozen Experiment 07 publication evidence: https://github.com/RhysWhite/branchsnv-validation/tree/2afa89f47815eac3b85b722c2fddd14d9c17f22e BRANCHSNV validation and analysis repository: https://github.com/RhysWhite/branchsnv-validation Associated BRANCHSNV software v0.2.0: https://doi.org/10.5281/zenodo.23518619 Licensing: MIT for software and CC BY 4.0 for research evidence and documentation, as specified in the supplementary archive.