Data & reproducibility
Current v10.2e derived outputs for 113 documents, 3,963 page units and 6,949 source-verified union candidates. Public files contain analytical outputs, paired blind human-review data and provenance—not API keys or source PDFs whose redistribution rights are unclear.
Current release — v10.2e
The two sealed provider runs contain 11,358 source-verified statements (OpenAI 5,675; Claude 5,683). Deterministic source-span matching and deduplication produce 6,949 candidates. The primary equal-vote ensemble retains provider disagreements and is accompanied by OpenAI-only, Claude-only and exact-code-consensus variants.
Validation status: calibration precision 0.980, recall 0.766 and span fidelity 1.000; recall failed the frozen threshold. Independent blind coding is complete for 365 candidates and 36 dual-empty units. Reviewer-specific design-weighted validity is 78.6% and 76.1%, inclusion agreement is 62.6% (AC1 = 0.426), and consensus adjudication remains work in progress.
Download the current outputs
DERIVED FILES ONLY · HASHED AT BUILD TIME · REPRODUCTION BUNDLE AND PROVENANCE ON GITHUB.
| File | What it contains | |
|---|---|---|
entity_scores.csv | Entity-level ensemble and sensitivity variants: purpose-centre scores, dominant purpose, privacy-family share and documented posture. | Download |
distributions.csv | Weighted code, family and ODR distributions by provider variant. | Download |
candidate_allocations.csv.gz | Compressed candidate-level ensemble allocations, including provider origin and fractional code votes. | Download |
analysis_summary.json | Candidate populations, provider overlap, exact-code/family/ODR agreement and dominant-centre counts. | Download |
human_validation_results.json | Blind pre-adjudication sample counts, design-weighted validity, intercoder metrics and dual-empty missed-claim yield. | Download |
macro_results.json | Measurement relation and institutional correlations for ensemble, provider and consensus variants with bootstrap intervals and permutation tests. | Download |
corpus_manifest.csv | Frozen 113-document inventory with filenames, archive members, page counts, source type and SHA-256 hashes. | Download |
jurisdiction_stage.csv | Current jurisdiction stages and verbatim evidence, including DCash 2.0 as cancelled after the February 2026 ECCB suspension. | Download |
codebook.csv | The current 35-label substantive codebook. | Download |
privacy_posture_v10_2e.csv | Website-ready ranking of exploratory privacy posture and privacy-family share. | Download |
deployment_manifest.json | Build provenance linking website headlines and downloads to exact local source hashes. | Download |
checksums.sha256 | SHA-256 checksums for every deployed file except the checksum file itself. | Download |
Canonical provider outputs, frozen prompts, provenance manifests, the completed pre-adjudication workbook, case-level human data and executable scripts live in the GitHub repository.
Reproducibility boundary
Deterministic: source verification, canonicalisation, candidate matching, ensemble allocation, scoring, figures and statistical outputs replay from frozen artifacts.
Provider-dependent: the original extraction requires the named model APIs. Raw responses are sealed before parsing so the published run can be audited without rerunning a model.
Human-dependent: the independent blind phase is complete and its reviewer-specific estimates are released. A single final consensus-validity estimate and provider-to-consensus agreement remain unavailable until the 319-case adjudication queue is resolved.
License & citation
Derived tables and website text are released under CC BY 4.0. Official source documents remain subject to their publishers’ terms.
@misc{bok_stroukal_cbdc_privacy_2026,
author = {Bok, Martin and Stroukal, Dominik},
title = {Institutional Determinants of Privacy Design in Central Bank Digital Currencies},
year = {2026},
note = {Working paper and v10.2e reproducibility dataset},
url = {https://cbdcdataset.org}
}