Data quality — provenance, coverage, freshness & completeness

Cairn's data quality is, before anything else, a provenance property: every figure traces back to an immutable official source, pinned by sha256 in an append-only manifest. This page reports that chain's integrity for each source — observable facts from the manifests, not a confidence score. It deliberately does not rate how "good" a number is; it shows whether the number is still pinned to the source it came from.

Configured sources

Pinned to immutable storage

Raw rows pinned by hash

8,707,444

What each status means

The pin status is derived deterministically from where a source's raw file is stored — it is a fact about the snapshot, not a judgement about the data:

pinned_r2 — the raw file sits in immutable object storage and is pinned by sha256. This is the auditable pin of record: the figure cannot drift without a manifest change and a failing build.
pinned_local — a file:// pin from an --offline run. Useful locally, but it is never the committed pin of record, so a pinned_local row in a published build is an integrity flag worth investigating.
unpinned — the source is configured but has no snapshot yet. It is shown honestly as unpinned, never filled in with a placeholder hash or row count.

Current pin of record per source

No Results

The hash shown is the claim; CI verifies it. This page reports the SHA256 each source is pinned to. A weekly reproducibility job re-downloads each raw file, recomputes its SHA256, and rebuilds the warehouse from it — so a silently mutated source file is a failing build, not a number that quietly drifts. See Architecture → for the full CI rail and Methodology & sources → for per-figure provenance.

How complete is the coverage?

Provenance answers "is the number still pinned?". Coverage answers a different question: how much of each source's official aggregate does Cairn's mart actually capture, and how much of the CBS national total can be attributed to a NACE sector? These are the same reconciliation figures the CI benchmark already computes — assert_national_total_reconciles (CBS) and assert_euets_coverage_within_eea (EU ETS) — surfaced here as standing facts rather than discarded once the test goes green.

No Results

Descriptive, not a score. covered_share describes how completely each source's official aggregate is captured, and — for CBS — unmapped_share describes how much of the national total cannot be attributed to a single NACE section. Neither is a confidence or quality score on any individual figure. reconciliation_drift near zero confirms the mart's own total still matches the official aggregate it derives from.

How current is the data?

Provenance and coverage answer "is it pinned?" and "how much does it capture?". This section answers a third question: how current is it? For each source, the pinned release and ingest date, the latest calendar year its benchmark mart actually covers, and the observed lag between "now" (this page's own build time) and those two facts.

No Results

Descriptive, not an SLA or alarm. ingest_age_days and coverage_lag_years are measured against this page's own build time, so they drift with every rebuild — they are not pinned figures. latest_covered_year is NULL for a source with no benchmark mart yet. No "expected next release" date is ever computed: several sources publish on no fixed cadence, so only the observed lag is shown, never a predicted one. Watching for an actually new upstream release is the weekly reproducibility job's and the no-LLM dispatcher's job (scripts/check_freshness.py), not this page.

How complete are the nullable columns?

Provenance, coverage, and freshness answer "is it pinned?", "how much does it capture?", and "how current is it?". This section answers a fourth question: for columns that are deliberately nullable, how fully are they populated? lei, allocated_total_t_co2eq, and surrendered_allowances_t_co2eq on the two EU ETS marts, and the UNMAPPED sentinel NACE on the CBS sector mart, are each legitimately NULL (or UNMAPPED) some of the time — not-yet-mapped in a reviewed seed, or genuinely absent upstream. This tracks that completeness as an observable fact, letting reviewed-seed coverage (e.g. the LEI mapping) be watched as it grows over time.

No Results

Counts only, never a fix. A NULL (or UNMAPPED) value here is a legitimate, documented state, not an error — this page never imputes or fills one in, and populated_share is never a quality verdict on the underlying emissions figures themselves. The tracked columns are a curated list, not every nullable column in every mart.