Varion benchmark outputs: callsets and hap.py scoring for the SNP caller R2 and the indel caller nindels / R3.1 on GIAB HG004 and HG007 subsets

Generated outputs behind the benchmark tables of the Varion project (software: https://github.com/krmox/AI_DNA_ANALYZER, release 3.1.0, doi:10.5281/zenodo.23856530). Contents: variant callsets (indel caller nindels, configurations A and A+B, and the depth-dependent front end R3.1; SNP caller R2; bcftools comparators), normalised query VCFs, hap.py 0.3.15 (xcmp) scoring output (summary, extended, metrics and ROC tables, logs) for 298 reanalysis runs and for the R3.1 development and test matrix, development logs, the R3.1 development report and partial test report, and analysis scripts. 5,657 files in 6 archives (1.3 GB), plus documentation. Cells with files in this record: HG004 chr20 and chr21, HG007 chr17 and chr19 (GRCh38; GIAB v4.2.1 truth and confident regions; Illumina short reads), downsampled to 10-100x with one seed per depth. The SNP and indel callers are evaluated separately; a result for one says nothing about the other. Not included: (1) the public input data (GIAB reads, reference, truth sets); checksums and sources are in EXCLUDED_INPUTS.csv, and the official GIAB/NIST releases are at https://ftp-trace.ncbi.nlm.nih.gov/ReferenceSamples/giab/ and https://www.nist.gov/programs-projects/genome-bottle; (2) GIAB truth records: the hap.py annotated VCFs and truth subsets (1,234 files) were left out; (3) HG001-derived files (3,183 files: all files under HG001 cells, and the reports, tables and run logs that carry HG001 results or identify HG001 runs), left out for now because redistribution of data derived from the pilot genome was not established; for the same reason this record contains no summary tables. Every file left out is listed in EXCLUDED_FROM_DEPOSIT.csv; unavailable sources are listed in LIMITATIONS_AND_GAPS.md. MANIFEST.csv lists the sha256 of every file and verify_deposit.py checks them. Two transformations were applied to the files: the key 'environment' (a dump of the process environment) was removed from hap.py runinfo.json files, and machine-specific absolute paths and the host name were replaced by documented symbolic roots (VCF records are unchanged). See SANITIZATION.md. Rights: the original generated outputs and the documentation are offered under CC BY 4.0. The analysis scripts are GPL-3.0-only. No GIAB truth records and no HG001-derived files are part of this record, and no licence is claimed for them. See RIGHTS_AND_LICENSES.md. Known limitations: several native BAMs and fetch logs are no longer available, one binary is tied to its source only by output equality on one cell, deposited indel VCFs come from a development build that differs in VCF representation (not in calls) from software release 3.1.0, and the frozen R3.1 switching thresholds do not follow their pre-registered protocol. The record does not contain the original HG001 chr19 validation runs, the HG004 chr20 regression runs, the SNP replication campaign and DeepVariant files, the HG006 chr16 and HG003 chr8 records, or the scripts that assemble the manuscript tables. The tables of the associated manuscript cannot be regenerated from this record alone; the per-run hap.py summary files it contains allow the cell-level metrics of the HG004 and HG007 runs to be recomputed. See LIMITATIONS_AND_GAPS.md and SOFTWARE_PROVENANCE.md before reuse.