Method
A summary for readers. The authoritative description is the project brief and the decision records in the repository.
The question
When the makers of a retail CBDC say it will respect privacy, what exactly do they say, and what have they committed to?
Which projects
44 projects: 43 national retail CBDC projects and the digital euro. A currency union is one project; documents of the ECB and of the national central banks of the euro area belong to the digital euro.
Which documents count
Only official documents of the body competent to decide for the project, on its official channels. Reports of the BIS, the IMF, consultancies, trackers and the media are never evidence. Each document collected is decided by the authors as one of:
- Evidence: read by the measurement.
- Companion: cited for the project's status, timeline or legal state (for example a press release announcing a report, or an act of parliament), never measured. A binding text issued by the central bank itself stays evidence.
- Not used or wholesale only.
Rule: SOURCE_RULE_D1. The search for missing documents was run for every project with the same rule.
Two layers, never mixed
Talk. Every statement in a project's own documents about users' privacy, identifiability or access to their data, split into sentence units before any model sees them and labelled by aspect and voice.
Commitment. Twelve design features (for example: can the central bank link a payment to a person?), counted only when the competent body has decided them or the system runs them, each with a located quote. Silence is never read as a commitment.
Languages
Documents not in English are read in the issuer's own English edition where one exists, otherwise in a machine translation made through a provider's API with a full run record. Machine translation is used only for projects with five or more evidence publications. Its quality is being judged by both authors on a benchmark of five documents that also exist in the issuer's own English; how far translated text may be used follows from that judgement.
Checks
Every model run leaves a run record (exact model, prompt hash, inputs, per-call log). Human coders code before they see model output. The instrument has been audited four times by an independent model; the open findings are listed in the repository. Nothing is reported as validated until the authors' agreement meets the thresholds set in advance.
What the dataset will not claim
No ranking of projects by how much they talk about privacy, no "silent" projects, no score combining talk and commitment, and no causal language.