Worked answers

Read after you have decided. Where more than one reading remains possible, it says so.

  • cache row: one saved lookup: a printed place name and the identifier it was matched to.
  • manifest join, containment: matching printed names against a table of historical polities; accepting a match because one name occurs inside another.
  • ledger, dossier: the file of merge decisions with reasons; the evidence laid out for one cluster.

All terms: glossary.

Start with R1, R3 and R8 after you have decided. The other cases are further practice. These are the preparer’s answers with the evidence; each decision is for the proposal in front of you.

R1 (Structure). Blakely’s record does not stand

Reject. The entry prints no birth clause. Its only date, “Apl., 1898”, is the appointment, and it went into birth_year while the event lost its year. The prompt forbids exactly this inference. Also “off.” became “officer”. This record contributed one of eight major findings across six records in the 186-record audit; the category is HALLUCINATION. The fix is a guard: a birth year may only come from a printed “B.” clause, and an event with no year while the entry has an unassigned date is flagged.

R3 (Clean). Abery and Carbery are one person

Accept the merge. The rules allow a merge on an OCR surname variant when the forename and career are consistent and one identical dated event is shared. Here: same forename, same qualification (M.B.C.M.), the same 1867 Ceylon appointment, and ABERY attested only in the 1897 edition where CARBERY also appears. The ledger records the reason: “ABERY is OCR garble of CARBERY”. A reviewer who wanted more could check the 1897 page image for the lost characters.

R8 (Ground). Harris can be linked without settling his birth year

Accept with the disagreement recorded. The retrieved candidate shares the full name, Christ’s College and the Newfoundland governorship. Read the biography as a whole, allowing for OCR errors: an isolated birth-year discrepancy should not split an otherwise matching career. Link the identity and review the disputed date separately. Leave the identity unresolved when the wider biography does not supply enough evidence to decide.

The proposal keeps birth_year_as_printed: 1858, records the external value 1855 separately, and flags the conflict. Accepting the link does not authorise changing the source field. The OCR could be wrong, the printed book could be wrong, or the external record could be wrong. The excerpt alone cannot decide which.

Next checks: inspect the 1927 page image, compare another edition, and trace the external date to its cited source. Do not count repeated copies of the same account as independent corroboration. Retrieved evidence.

Further practice

R2 (Structure). Balmer’s “do.” is not a place

Reject as it stands; accept once resolved. The clause names no place; the rule says inherit the previous one, which is the Kenya-Uganda-Tanganyika service. The model copied the abbreviation. Because this error is systematic across the sample, the fix is a deterministic forward-fill: any place equal to “do.” or “ditto” inherits the previous event’s place, with place_inherited set. No model rerun needed.

R4 (Clean). A’Beckett and Beckett are two people

Reject the merge. The cluster exists only because of the suffix key and the forename. Birth years are 62 years apart, and the conflicting careers support keeping them separate; the year discrepancy is not the sole reason. The careers share nothing: Lincoln’s Inn and the Victorian bench against Nigerian telegraphs. This is the normal case in the cluster sample: the clustering is deliberately loose so that real variants meet, and most clusters resolve to “keep apart” in seconds.

R5 (Ground). “Union of S. Africa” is not Q5155572

Reject. Q5155572 is the British South Africa Company’s territory, seated at Harare. Q193619, the Union of South Africa, is what the surface names. The error came from a containment match on the string “South Africa” during the manifest join; it affected 960 Colonial Office events and was visible on the atlas as a Rhodesia node with far too many arcs. The fix patched five cache rows with a stated reason, and added a containment guard to the join. The October rebuild added the date-aware check: pre-1910 records cannot use the Union, explicit towns are retained, and unresolved jurisdictions are not replaced with a guess.

R6 (Ground). Brewster’s “Victoria” is British Columbia

Reject the proposed grounding; ground to British Columbia for this person. Every other clause in the entry is British Columbia, and Victoria is its capital and an electoral district. A surface-keyed cache cannot know that; a per-person resolver can. That repairs the geography only. The saved graph still turns the 1912 defeat into an appointment lasting to 1916; the source does not support that office held. Keep or type the defeat separately, and flag the appointment assertion for correction.

The place fix routed “Victoria” through a context rule (British Columbia signal in the career → Q1973, which geolocates to its seat; otherwise the Australian colony). Fixing it exposed a second bug: per-person overrides had never applied because the resolver and the graph keyed people differently. Both corpora were re-emitted.

What remains open: six “Victoria” surfaces (“Victoria district”, “Victoria county”, the Cape variants) mean different places for different people and are left ungrounded by design until each is resolved per person.

R7 (Ground, extra). Price is Q123279571, on the name and dates

Accept, with the evidence recorded. The item has a rare name with the alias covering the spelling, birth and death years consistent with the career (scholarship 1875, writership 1878), and a Tamil description meaning Sri Lankan politician, which fits three terms as mayor of Colombo. It has no career statements, so the match rests on name, dates and place of activity, and the record should say so. The pipeline’s conservative matching gate rejected it because the name spelling differs; that is the gate doing its job, and this is the kind of case a human review queue exists for. A reasonable person could also mark it unresolved pending a Ceylon directory; what they cannot do is treat the gate’s rejection as evidence that no item exists.

The pattern

Across these cases the decisive evidence was never the model’s confidence and never the identifier’s existence. It was the source’s own words (“Apl., 1898” is an appointment; “do.” is a ditto), the dossier’s dates and careers, the item’s statements read rather than assumed, and the map. A correction needs a reason, a rerun and a check of other affected records. The ditto replay demonstrates that on a small sample; the Wodehouse and Brewster repairs show the same loop in the public atlas, including a wider check of affected records.