The dataset
The dataset is a set of CSV tables, one determination per row, with every derived value traced
to its source. The tables join on record_id (Chapple corpus) or report_id + lab code (the
extension).
The tables
| Table | Rows | What |
|---|---|---|
dates.csv | 10,665 | The Chapple corpus, typed and QA’d — raw BP, error, lab code (validated), coordinates, δ¹³C, notes, QA flags |
dates_extension.csv | 969 | Post-2019 determinations (excavations.ie, RIA, TII) with δ¹³C/δ¹⁵N |
calibration.csv | 11,634 | Per-record reservoir class + calibration summary, both corpus and extension |
calibrated_intervals.csv | 54,287 | One row per IntCal20 HPD interval — multimodality preserved |
site_type_map.csv | 1,508 | The controlled site-type vocabulary: raw string → ASI class + period |
material_inferred.csv | 10,665 | Sample material recovered from Chapple’s notes |
reliability_inferred.csv | 10,665 | Chapple’s own “don’t use” / “anomalous” judgements, recovered from prose |
xronos_enrichment.csv | 10,289 | Cross-validation + δ¹³C enrichment against the compilation ecosystem |
Download the current tables from the repository: git.sr.ht/~calgacus/radio-rian → data/derived. The full data dictionary is docs/data-dictionary.md.
Two principles worth knowing before you use it
The raw determination is the crown jewel. The calibrated intervals are a convenience built on top and labelled as such. Serious analysis recalibrates from raw BP ± error; what it cannot reconstruct is the metadata that makes calibration sound — which is what this project recovers.
Nothing overwrites the source. A correction adds a flag and a reason; it never changes Chapple’s published value. Where two sources disagree, both are recorded and the conflict is a field. The dataset does not adjudicate.
decided_by=rule — machine first passes awaiting
human review. Use the raw determinations freely; treat the derived judgements as drafts.