Dated read-only audit: 26 September 2026, 17:23–17:30 UTC; earlier cohort measurements identified below. These are recorded observations, not live status checks. Current health responses do not automatically resolve these findings, and the interface does not repair or reprocess data.
For 1 September 2026 UTC, 427 stored rows were compared with 384 current USGS preferred event IDs: 375 earthquakes, six quarry blasts and three explosions. All 384 were present locally, but nine native non-earthquakes were classified as earthquakes. Thirty-two additional local rows matched provider aliases; eleven local IDs remained unresolved. This bounded comparison does not establish whole-archive completeness.
The audit identified 13 network uu / type mh values of −9.99 as documented uncomputed-magnitude markers, dated 2017–2019. Preserve the native value and lineage while excluding this exact source-specific marker from numeric magnitude statistics. Negative magnitudes in general remain valid; they must not all be discarded.
The configured guard was 160 GiB and the database measured approximately 160.86 GiB. The watcher was restarting after the guard rejected further work; the filesystem still had free capacity. This was not a demonstrated full-disk incident. A capacity and retention decision is required; the dashboard does not disable the guard.
The 17:02 UTC poll partly succeeded but rejected revisions for 2026-09-23T16:10:00-RBE-001 and 2026-09-24T08:09:00-HSS-001 with source_version_collision. This is not established to be a provider network outage. Conflicting observations need lineage-preserving adjudication; retrying does not resolve scientific authority.
GVP catalogue and eruption-history refreshes reported provider_network_error, while weekly activity remained operational. A blanket green volcano status would conceal those differences. The underlying network cause was not established in this audit.
Only five of 26 registered sources had rows in core.dataset_coverage, covering 855 stored partitions. Their complete labels are method-specific claims, not fresh independent provider recounts. Dedicated scientific controllers can ingest while ordinary scheduler flags are disabled, and older completed partitions must not be treated as live sensors.
The June–26 September query matched 46,547 stored records, including 4,863 mb records represented by 598 buckets. Other magnitude types are separate scales, not absent events. The all-earthquake frequency view combines record counts, never unlike magnitude values. These are dated query counts, not live or worldwide completeness estimates.
The previous example grid had no stored rows. The verified sample power:+35.000:-115.000 returned both daily rows for 1–3 September, but solar radiation was missing in both. Missing numeric values remain unknown, not zero. Old failed NOAA windows also overlap newer recorded coverage, so an old failed job alone does not justify replay.
Set the query in the rail and search. The atlas, graph and records all describe that one selection.
Provider-attributed information. Coverage and scientific quality vary. Not an emergency-warning service.