The map finds the errors

Minutes 40–46. Three errors seen in the atlas, what caused them, and what fixing them changed.

  • cache row: one saved lookup: a printed place name and the identifier it was matched to.
  • surface: a place name exactly as printed.
  • manifest join: matching printed names against a prepared table of historical polities.
  • containment match: accepting a match because one name occurs inside another; the cause of the South Africa error.
  • canonical: the one agreed entry for a thing, to which duplicates are folded.
  • re-emit: write the graph out again after the fix.
  • stratum: one slice of a sample, chosen to oversample the risky cases.
  • false negative: a link or merge that should have been made and was not; invisible on a map.

All terms: glossary.

A wrong grounding is invisible in a table of 180,000 rows. On a map it is a dot in the wrong hemisphere. The project keeps a running list of defects surfaced while building the atlas, and a review of 3 September 2026 that re-fetched every QID in the graph and loaded the atlas headlessly to look for more. The examples below follow the errors through to their published repairs. Saved outputs show the earlier state; the rebuilt atlas shows the result.

A premier of British Columbia in Australia

Harlan Brewster’s 1918 entry (saved): newspaper work in New Brunswick and Boston, removed to British Columbia in 1897, elected for Alberni, “defeated by Sir Richard McBride, in Victoria, 1912”, re-elected at a by-election in 1916, premier of British Columbia 1916. The saved graph gave him two events as member for Victoria. The bare surface “Victoria” was grounded to the Australian colony, Q56850459, by the surface-keyed cache. The atlas showed a British Columbia premier with a seat in Melbourne.

Seen on the map. Cause: a cache keyed on the surface string, with “Victoria” resolved once for everyone. Fix: a per-person context resolver. If the official’s other postings signal British Columbia, “Victoria” is British Columbia (Q1973), which geolocates to its seat; otherwise the colony. Also found while fixing it: per-person overrides had never applied at all, because the resolver keyed people one way and the graph another. Repairing that fixed a second, unrelated split between the North-Western Provinces of India and of Ceylon. Both corpora re-emitted.

A plausible map can still be wrong

Now read the source again: Brewster was defeated in Victoria in 1912. In the saved graph, that clause became a legislative assembly posting for Victoria from 1912 to 1916. Correcting the place leaves an unsupported appointment on a plausible map.

Pause and decide: does this record stand after the geography is fixed? What must be preserved, and what must be removed from a table of offices held?

The rebuilt graph now preserves the defeat as an event of its own type and removes the office-held assertion. The public atlas lists it under other recorded events. The same check applies to explicit electoral-defeat events elsewhere in the corpus; military victories and resignations require different readings. A checked location is not a checked career.

The Union of South Africa at Harare

Five surfaces, “Union of S. Africa”, “S. Africa”, “S.A.” and two more, had been joined to the Wikidata item for the British South Africa Company’s territory, Q5155572, because the join matched “South Africa” by containment. 960 Colonial Office events, 283 arcs in and 289 out, plotted at Harare.

Seen in the September review, on the map, as a node with far too many arcs for Rhodesia. Cause: a containment match in the manifest join. Fix: five cache rows patched to Q193619 Union of South Africa and a containment guard added to the join. That first repair still treated pre-1910 “S. Africa” too broadly; the October date checks completed this part of the correction. The patched rows carry the reason (cache rows).

The 7 October audit found 1,120 pre-1910 Union assignments across both Lists. The repair removes the anachronistic jurisdiction, retains explicit towns where available, and leaves ambiguous cases unresolved. It also removes the nearest-posting fallback and selects seats by date. See the audit and review queue.

Two Baluchistans

The colony layer had Baluchistan under Q843, which is Pakistan, for fifty India Office events, and a second Baluchistan node under a different QID. The atlas drew two nodes with the same label. Fix: one canonical entry in the colony table, the duplicate folded in.

What a fix changes

A source-specific correction records its evidence; a shared rule is checked against everyone it affects. After each one the pipeline re-emits both corpora and the atlas is rebuilt. The review then checks two things: the cases the fix was meant to repair, and the previously accepted cases that changed. The second group is where new errors hide.

The same discipline for structure: the every-tenth audit’s error catalogue became a post-processing pass with tests, and the three name probes that found the dedup defects are regression tests. The probe that found the error is the test that keeps it out.

The October checks also tested the actual lines drawn in the browser. Cameron’s temporary Windward governorship remains on the map and connects to Gambia. His Ottawa conference visit remains a written event without creating a map stop. A missing line could otherwise make a sound appointment look as though it had been removed.

What the map cannot find

False negatives. An official whose “Victoria” was left ungrounded does not appear anywhere wrong; he simply has no arc. A merge that should have happened leaves two short careers instead of one long one, both plausible. For those the project built the silver standards: 200 hand-labelled cases, stratified so that the risky strata (rare surname, overlapping eras) are oversampled, scored against the machine output. On the Colonial Office List never-biographied population the judge’s “different” verdicts in the rare-surname stratum turned out to have a false-negative rate of 76 per cent, which no map would have shown.

Now: the exercise. Work through R1, R3 and the new R8. Give each decision a reason; for an unresolved case, say what evidence you would seek. The other five cases are further practice.