The provenance archive

Transcribed country observations

Retired 2026-08-24 · 1,671 identifiers

The original country layer carried 1,671 hand-transcribed values marked as awaiting automated collection. World Bank data replaced them with full annual series under CC BY 4.0. Keeping both would have meant publishing two competing figures for the same fact. The transcribed identifiers no longer resolve.

Hand-transcribed rainfall normals

Retired 2026-08-24 · 5 identifiers

Five UK 1991–2020 rainfall normals were transcribed before automated collection. They are now computed from the full 190-year series and differ by up to 3.8% — the England normal moved from 838 mm to 870 mm. The computed values supersede them.

Finnish daily discharge, while reorganising how the Finnish record is stored

What was wrong. Two Finnish basins had grown larger than a single page can afford to read, so the archive was collected again to split them on Syke’s own sub-basin boundaries. That run happened during a network outage. 163 of 671 stations could not be fetched, and because the existing files had been cleared first, there was nothing left for the usual protection to carry forward. The Finnish record fell from 10,327,195 values across 671 stations to 7,647,178 across 508 for the length of one afternoon.

How it was found. By the run report itself, which stated plainly that 163 stations could not be fetched and that 508 of 711 were stored. The counts were checked against the previous run before anything was rebuilt on top of them.

What was done. Collected again from Syke once the connection returned, this time keeping what was already held, so only what was fetched was replaced. Nothing was published from the incomplete state: the fault was caught before anything was rebuilt on top of it. The lesson is about the instruction given rather than the code — the protection that carries a record through a network failure exists precisely for this, and clearing the files first removed it. Recorded because a near miss caused by an operator deserves the same honesty as a defect.

Dutch water level at tidal stations, when Rijkswaterstaat was first collected

What was wrong. Rijkswaterstaat returns several different series for one station in a single response, not several pages of one series. At tide gauges the same timestamps arrive twice — once measured by an instrument, once predicted by harmonic analysis — differing by tens of centimetres. Nothing in the values, the unit, the reference plane or the quality codes tells them apart; only a process-type field does, and Hydrionis was not reading it. Roughly half the daily means at affected stations were computed from predicted tide and published as measured water level.

How it was found. By the collection run reporting 5,585,977 conflicting duplicate timestamps. A station with one instrument had reported 8. The figure was in the run report before the data was published and was not questioned until afterwards.

What was done. The whole Dutch record was withdrawn rather than patched, because the affected stations could not be separated from the sound ones after the fact. Hydrionis now separates response blocks by process type, reference plane and instrument before removing any duplicates, rejects predicted series outright, and refuses levels measured against a datum other than NAP. The data was collected again from scratch: the corrected run rejected 6,312,728 predicted readings and 1,838,461 measured against a datum other than NAP, and the conflicting-timestamp count that revealed the fault fell from 5,585,977 to 2,496. What is published now is 111,725 daily means across 443 series, against the 126,569 first published.