Tegiszhol genome-grave links: kurgan- and grave-level provenance for published steppe genomes

Grave-level provenance for published ancient steppe genomes, and a per-grave join to the Tegiszhol radiocarbon register. Archaeogenetic compilations (AADR, Poseidon) record each ancient genome's site, culture, coordinates and date, but not which kurgan or grave it came from — a check of Poseidon's 51 .janno columns finds no kurgan/grave/burial field at all. That context does exist, scattered across ten studies' supplementary files in ten incompatible formats, two of which are not machine-readable without OCR. This dataset harvests it into one table, keyed by genetic ID, with a verbatim source snippet and pointer for every record. Contents genome_grave_table.csv — 391 published genomes placed to a specific kurgan and grave, from 10 studies (Narasimhan 2019, Jeong 2020, Saag 2021, Damgaard 2018, Jarve 2019, Krzewinska 2018, Wang 2019, Kilinc 2021, Allentoft 2015, Mathieson 2015). 276 carry an AADR laboratory code; 23 cross-link to an independently extracted Russian-language radiocarbon determination. grave_spine.csv — 817 graves across 522 sites, joining the radiocarbon register to the genome table one row per burial: 23 linked (a date and a genome), 447 radiocarbon-only, 347 genome-only. GENOME_PROVENANCE.md, GRAVE_SPINE.md — datasheets. verification_score.json — full blind-verification results, including every disagreement. Measured accuracy A blind verification round on Prolific: six independent annotators, each transcribing the same stratified 48-record sample from red-boxed crops of the primary sources, never shown the pipeline's answer. fieldpipeline vs human consensushuman-human agreement (ceiling) kurgan100% (48/48)69% grave94% (45/48) — 100% after source adjudication83% burial phase100% (48/48)94% skeletonnot reportable17% All three grave disagreements were adjudicated against the source, and in every case the majority of annotators — not the pipeline — was wrong: DA229 has no grave number in its source (annotators copied the skeleton number), FNO003 reads pogrebenie 4 (Russian for grave, unknown to English-screened annotators), and RISE392 reads "burial B", a letter grave the pipeline originally dropped — a genuine defect, since fixed. Consensus is therefore reported but never treated as ground truth. The skeleton field is flagged low-confidence and should not be relied on. Its 17% inter-annotator agreement reflects an instrument defect (one annotator systematically entered the sample ID), not a measured property of the data. Caveats Provisional. Cross-script joins are by laboratory code only; site names are never fuzzy-matched between Cyrillic and Latin, because a false genome-grave link is worse than none. genome_only therefore means "no date we hold", not "no date exists". Central-vs-intrusive burial phase is populated only where a source states it (7 records). Most sources do not, and it is never inferred. Two sources (Allentoft, Mathieson) were extracted with a benchmarked vision/language model rather than a deterministic parser, and are weighted accordingly. Attribution Every record points at two prior works: the excavators who dug and dated the grave, and the geneticists who published the genome. Please cite both. This dataset is a pointer to their work, not a replacement for it. Archaeological cultures are recorded exactly as the source states them and are never equated with ethnic groups or modern nations.