A standard for auditable reference states: what one is, the six states a query can exit through, the six requirements a conformant collection must meet, and the boundary that keeps it honest. Ours. Disclosed to clients in confidence.Verified
A government commissioning a view of its own economy is usually offered a dashboard. A dashboard is a display surface: it shows numbers, and it is silent about what stands behind them.
When a minister is asked, in public, why a tile said what it said — which source, observed when, coded by whom, checked how, and superseded by what — the dashboard has no answer, because nothing in its construction required one. The failure that follows is never "bad data" in the abstract. It is always a specific sentence, in a specific published surface, that the evidence no longer supports — discovered by someone else first.
A reference state is built so that question can always be answered. Every claim carries its evidence, its dates, its checker and its status. A dashboard may sit on top of it — but every tile inherits an answer-state from the record beneath, and shows it as plainly as its value.Verified
Ask a reference state anything and it returns two things: an answer, and a declaration of what kind of answer it is. "Cannot answer" stops being a conversational shrug and becomes a governed output. Six states. Set them yourself.
Conformance is stated for a declared scope, never as a blanket. A conformant collection has to be able to answer six questions about itself. The questions are here. What counts as a passing answer is disclosed under engagement.
The taxonomy, the requirements, the collection architecture and the implementation that codes to them are Nixus's own. They are shown to clients in confidence, and what we commit to under them — conformance, for a declared scope — is committed in writing under the engagement.Verified
A conformant collection is built from three layers of evidence, each with its own provenance and verification discipline, and separates decision-grade readings from restatements of the input by testing every measure against a null. Readings produced under it are demonstrated, not published.
Superseded versions are retained, not overwritten.