A 250-jurisdiction AI-observability index (dataset)

A frozen, version-controlled dataset that scores public AI-governance observability for 250 countries and jurisdictions (freeze date 2026-06-01). Observability here means whether public records reveal the governance arrangements that would let a competent authority, public buyer, auditor, court, citizen, or researcher trace AI-system and AI-agent use from legal basis through to lifecycle evidence. It is not a compliance rating and not a count of deployed agents. The instrument has ten indicators (I01-I10), each scored on a 0-3 public-evidence scale and weighted to a 0-100 composite, with five subindices and a confidence-adjusted composite. The distinguishing feature is a per-cell evidence ledger: every score traces to logged public sources, each carrying a recorded confidence flag (H/M/L). The unit of the dataset is the scored cell: 250 jurisdictions x 10 indicators is roughly 2,500 ordinally scored, individually evidenced observations. World mean raw = 41.2. The EU-27 subset has a raw mean of 59.4/100 (classes A=3, B=13, C=8, D=3). A blind second coder re-scored a stratified eight-jurisdiction sub-sample (80 paired ratings): Krippendorff's alpha = 0.747. Convergent validity against the World Bank GovTech Maturity Index (GTMI, 2022) is Spearman rho = 0.72 globally (n = 198). Contents: the per-country index, a leaner numeric matrix, the per-indicator matrix with confidence flags, the four-bin per-source evidence ledger, per-country narrative profiles, the frozen codebook, a machine-readable schema, and a freeze manifest with SHA-256 hashes. See README.md. The accompanying data descriptor documents what is in the data, how its integrity was checked, and what it can be reused for. The related research article (preprint DOI 10.5281/zenodo.20491780) builds a theoretical claim on these data and is a separate output.