Corrections

This archive said

Twenty-four of this archive's 112 households named a De Franceschi woman as the father and her husband as the mother. «father: Michiela De Franceschi · mother: Domenico Giachin». Forty people's roster entries said things like «Domenico Giachin — named as mother on 2 baptisms».

What is true

Every one is inverted. The women are the mothers; the men are the fathers. All 24 are De Franceschi daughters married out — which is why the error clusters exactly where it does.

How it happened

The households built from the FamilySearch index take the two parent names the index returns and assign them by position, not by sex. Where the De Franceschi parent was listed first — which is what happens when the mother is the one carrying this archive's surname — the pair went in backwards. So the bug bit precisely the people this archive is worst at tracking. The whole Vodnjan argument is that «the surname keeps its sons and loses its daughters»; twenty-four married-out daughters were in the data and the data had them down as men. Found while testing something else entirely — chasing whether one Michiela De Franceschi fu Giuseppe was a daughter of the Giuseppe this archive already documents. The check that found it is now in `scripts/audit.py` and runs on every publish: a female given name in the father field. It would have caught this the day it was introduced.

Corrected in place, here →

This archive said

That FamilySearch's full-text search returned 1,952 results for Gologorica, 1,719 for Gračišće and 1,885 for Defranceschi — and, noting how implausibly alike those three numbers were, that they «have not been tested» and should not be trusted.

What is true

Every one of the three is wrong, and the real figures are not close. Gologorica returns 19. Gračišće three pages. Crikvenica fifteen. Ližnjan two. Defranceschi returns about 9,900 — five times what was published, not a fifth of it. And Vodnjan returns nothing at all: the Croatian name post-dates every record in the corpus, which uses Dignano — about 12,200 of those, most of them the other Dignano, in Friuli.

How it happened

The number was scraped with a loose pattern — `([0-9,]+)\sResults` — off a page whose visible text ends «… Border Crossing Records 1924-1952\nResults per page:»*. It matched a year out of a collection title. «1,952 results» was a border-crossing collection's end date standing next to the words «Results per page». What makes this worse than a scraping bug: the page said the numbers looked wrong and published them anyway. Three unrelated searches returning numbers of the same size was called «the shape of a number that counts something other than what you think» — and then the one check that would have settled it, look at how many results are actually on the page, was not run. It takes one request. Flagging a doubt is not the same as resolving it. A figure you have publicly called untrustworthy should not be in the sentence at all. The finding underneath is unharmed and is better than it looked: 19 hits for Gologorica is a set small enough to read end to end, which is exactly what makes the village sweep worth doing.

Corrected in place, here →

This archive said

Nothing — that is the problem. The account of Paputchka's death, some five hundred words, silently vanished from the transcription.

What is true

The scanner destroyed its opening sentence, returning «WS CAPES FUG PS PEPE AU SII Pony FRU. OI WC PI PO UT UU TET» where she had written «We experienced "true" love with our Siberian Husky Paputchka». A rule I wrote to strip scanner noise matched that line — and because prose is stored one line per paragraph, deleting the line deleted the entire paragraph.

How it happened

Restored word for word from the original, supplied by David, and marked on the page as restored so no reader has to wonder which text they are looking at. The lesson is specific and unpleasant: a cleanup rule that operates on a whole paragraph can destroy a whole paragraph. Noise-stripping must be scoped to the noise — a few tokens — and never to the container it sits in. It was also found by the reader, not by me: nothing in the build warned that five hundred words had gone, because nothing was counting. A reconciliation was then run over the whole page — every substantial line of the raw scan tested against the published text — and it found four more losses of the same kind: the closing paragraph of the husky piece, which had been left beginning mid-sentence at «anger and rage»; a sentence from A Suicide Avenged; a line of dialogue from it; and part of the godparents letter. All four are restored and marked. One passage is unrecoverable — three lines describing a person, from the 2004 course — and stands on the page as a stated gap rather than a silent one.

Corrected in place, here →

This archive said

Her own introduction to the first poem was published word for word, and it named the man she was then married to in a sentence describing «a long standing affair, committing adultery for years».

What is true

That should never have gone up. It is an allegation of adultery against a named living person, published by this archive. That the words are hers and that she published them herself does not transfer the risk or the discretion — the archive chose to reproduce them.

How it happened

The name is removed; the sentence now reads «my marriage was over». It is the single alteration to her text anywhere on that page, and the page says so. The failure was not ignorance of the rule. The same rule had been applied an hour earlier, and correctly, to a different document — a statement of allegations that was excluded entirely for naming living people. Then a transcription was published verbatim without running the same test over it. A rule applied once is not a rule; it has to be run against every page that goes out.

Corrected in place, here →

This archive said

Two pages give her birth as 8 March 1926. A third page quotes the certified extract of the Senj birth register, attached to her son's 1951 entry, which gives 3 August 1926.

What is true

Both have stood on this site at once and nobody noticed. Written as numbers they are 8/3/1926 and 3/8/1926 — a day and a month exchanged, almost certainly where a European date met an English-speaking translator. This archive does not know which is right.

How it happened

Neither has been quietly deleted. Both now stand, marked as in conflict, and the Senj birth register for 1926 is an errand. The lesson is one this project keeps relearning: a date that appears twice in an archive should be diffed, not trusted twice.

Corrected in place, here →

This archive said

Anton Rudolf Defranceski, born 20 January 1890 at Klenovica house 22, died in 1943 — recorded here «per the family». The same page also has him on the Pfullingen list of foreigners in 1947 and certified by that town's mayor in 1950.

What is true

Those cannot both be about the same living man. Either the 1943 death is wrong — and it was only ever family testimony, never a register — or the Pfullingen documents are about a different Anton. The archive printed both and did not notice.

How it happened

Surfaced only because the family's oral history of 1948 made the widow's situation matter. Written up on Nineteen forty-eight, and the Crikvenica death register for 1943 is now an errand. Lore is worth collecting partly because it makes you re-read your own pages.

Corrected in place, here →

This archive said

That FamilySearch's tile server had blocked the death-register reading, recorded as `blocked` in the Search Register and written up on Method as a second instance of when the service says stop, stop.

What is true

The tile path was wrong. Tiles live at `…/image.xml`'s sibling `image_files/{level}/{col}_{row}.jpg`, and the reader was asking for `…/{level}/{col}_{row}.jpg` — no `image_files`. Today that same wrong path returns a plain 404; yesterday the security service answered the unknown path with a 403. So the reading was stopped by my own broken URL, dressed up as somebody else's refusal. A real block did come later, after several hundred tile requests, on `image.xml` itself — and that one is recorded honestly.

How it happened

Check that the request is well-formed before concluding you have been refused. A 403 on a path that does not exist is not a verdict on your conduct. The cheap test — does the correct URL work? — costs one request and was not run.

Corrected in place, here →

This archive said

A hand-tinted portrait of a bearded man in a bow tie, published as Josip Defranceschi, born at Gračišće 22 March 1863 — David's great-grandfather, and the only likeness the direct line had.

What is true

It is from the Imotski branch, and the sitter is a Josip of that family. No likeness of the Crikvenica Josip is known to this archive. The companion portrait published as his wife Ana is withdrawn too, because its only identification was that it matched the man beside it.

How it happened

Two branches' photographs sit in one family collection. A card carried the name, this line has a man of that name, and this archive matched the face to the name without asking which Josip — while checking every register plate on the same page against Pazin and Ledenice. The portrait was never checked against anything. Identified by David, who holds the originals. The retired 2018 GEDCOM spliced Imotski onto this line as well: the same seam has caught this archive twice.

Corrected in place, here →

This archive said

That the seventy-eight Vodnjan records were really Vižinada's, because resolving them to film 005497893 and asking the catalogue who made that film returned «Rimokatolička crkva. Župa Vižinada».

What is true

A reel can carry more than one parish. That one holds Vižinada's deaths 1754–1831 and Vodnjan's baptisms 1815–1866 (koha:963267). The records are Vodnjan's.

How it happened

The catalogue answers a film-number query with the first creator on the reel. The method was sound and the question put to it was wrong. It stood on the site for about an hour, published within twenty minutes of being reached and without a single image being opened. What made it look right: the one death on that reel really is Vižinada's — it is in the Liber Defunctorum of Castellier, Kaštelir.

Corrected in place, here →

This archive said

Eight reconstructed Vodnjan households and their children carried Vižinada as their parish in the roster, and the households themselves had no place at all.

What is true

They are Vodnjan. Eight households placed, forty-four rows moved, and Vodnjan moved out of unplaced into Southern Istria.

How it happened

The correction above was published on the narrative page and never reached the data. David asked why the headcount had not moved all day; it had not, because the roster was three commits behind the research and still carried the error the essay had already retracted. Fixed, and the rebuild order is now written into the queue notes.

Corrected in place, here →

This archive said

The Imotski family joined to David's own line, in the 2018 GEDCOM this archive inherited.

What is true

They are separate. The GEDCOM is retired — 8,286 people against the live tree's 15,643, and David's own verdict: «really old and really wrong».

How it happened

An old export was trusted as a source. It is now used for nothing. Note how often Imotski turns up in this list.

Corrected in place, here →

This archive said

Per-place record counts for this surname across sixty Istrian and Kvarner towns, and a research queue carrying «Vižinada — 53 records, unread», «Draguč — 25», «Dubrovnik — 38».

What is true

The place parameter on this search endpoint does not filter. Ten values were tested, including «Zzzznowhere» and «Atlantis», and every one returned the same total: 3,652. Vižinada and Draguč return identical result sets. Dubrovnik returns the unfiltered list. The queued «Vižinada 53» records are on the Vodnjan film and have already been read.

How it happened

The parameter is a relevance boost dressed as a filter — no error, no warning, just the whole set back in a different order. The sweep survives because its method independently checked each record's own place text before counting it; the queue did not, and carried three phantom tasks for days. The reliable route is to resolve records to their film and read the catalogue's item list — remembering that a film can carry more than one parish.

Corrected in place, here →