← Data Extraction

Page generated: 2026-08-28
Transcription as collected: unknown

Experimental transcription files used for phase B

This page shows the transcription used in condition B. It has errors and is not identical to the correct versions seen in the page viewer.

A model that faithfully copies an erroneous transcription is scored against ground truth that records the correct reading, and is marked wrong for ignoring the source image and replicating errors in the transcription. Here we record where models changed and answer from condition N to A to B in responses that reflect the intentional transcription errors.

Differences that change an extractable value

pageas collectedcorrectedwhat it affects
36todeschintodeschiniThe host's surname, three times on the page. In B the as-collected reading appears in 250 of 250 observations and the corrected one in none. N: none as-collected, 59 corrected, of 196. A: none as-collected, 85 corrected, of 211. Re-collected against the corrected text: none as-collected, 250 corrected, of 250.
39— absent —santi apostoliA parish line missing altogether, so the place of stay for the three travellers at that address lacked its location. In B the as-collected reading appears in 141 of 150 observations and the corrected one in none. N: none as-collected, 34 corrected, of 102. A: none as-collected, 44 corrected, of 112. Re-collected against the corrected text: none as-collected, 128 corrected, of 150.
36legnanolegnagoLegnago in the Veronese, not Legnano in Lombardy. This is the one case where re-collecting against the corrected text does not fix the output: the models write the Lombard town either way, so the transcription error and an independent substitution point the same direction. In B the as-collected reading appears in 50 of 50 observations and the corrected one in none. N: none as-collected, 27 corrected, of 28. A: none as-collected, 28 corrected, of 29. Re-collected against the corrected text: 50 as-collected, none corrected, of 50.
28tedescotodescoA traveller's surname. Note the models lean toward the erroneous form even with no transcription in front of them, so here the text reinforced a tendency rather than creating one. In B the as-collected reading appears in 28 of 38 observations, the corrected one in 10. N: 18 as-collected, 4 corrected, of 33. A: 16 as-collected, 10 corrected, of 33. Re-collected against the corrected text: none as-collected, 50 corrected, of 50.
28argièargiàA traveller's given name. In B the as-collected reading appears in 50 of 50 observations and the corrected one in none. N: none as-collected, none corrected, of 21. A: none as-collected, none corrected, of 26. Re-collected against the corrected text: none as-collected, 50 corrected, of 50.
28february 9february 10Not a spelling but a date: the day number was missing from the entry, so the arrival date of three travellers was unstated in the text the models were given. Without guidance the models omit the year entirely, which is why condition N carries neither reading. In B the as-collected reading appears in 150 of 150 observations and the corrected one in none. N: none as-collected, none corrected, of 126. A: 24 as-collected, 38 corrected, of 114. Re-collected against the corrected text: none as-collected, 150 corrected, of 150.

The rest of the diff does not change any extractable value. It is abbreviation rendering (Mons(ignor) → Mon(signo)r, Cav(alier) → Cav(alier)e, Mag(giore) → Mag(gio)r), diacritics (A dì → A di), verb forms in the departure formula (allog(giaron)o → allog(giaro)no), and on page 36 a large layout change in which runs of   were replaced by tabs.

What each condition produced for the affected values

Which reading each model wrote — not whether it scored. corrected form means the value carries the reading the corrected transcription and the ground truth agree on; transcription form means it carries what the as-collected transcription says instead; neither means it carries some third thing. Tested on the token, so a value that differs from ground truth in other respects still reports which form it holds.

This deliberately does not follow the score. §3 accepts todeschin for todeschini, so those answers are scored correct while still being the erroneous reading. Counted per matched traveller row rather than per run, so a value on five ground-truth rows contributes up to 50 observations per phase.

The last column marks the case this page exists to document: a model that wrote the correct reading in N or A and then followed the transcription into the error in B. 18 of 30 model-and-value combinations do that.

valuemodelNABN or A → B
Host surnamep36 · place_of_stay
todeschin for todeschini		claude50 neither50 neither50 transcription form—
claude_thinking5 corrected form · 42 neither9 corrected form · 37 neither50 transcription formfollowed the transcription
gpt40 neither36 neither50 transcription form—
gemini14 corrected form · 5 neither26 corrected form · 3 neither50 transcription formfollowed the transcription
gemini3140 corrected form50 corrected form50 transcription formfollowed the transcription
all models59 corrected · 137 neitherof 19685 corrected · 126 neitherof 211250 transcriptionof 250—
Parish linep39 · place_of_stay
absent for santi apostoli		claude1 corrected form · 19 neither20 neither30 transcription formfollowed the transcription
claude_thinking3 corrected form · 19 neither2 corrected form · 20 neither24 transcription form · 6 neitherfollowed the transcription
gpt10 neither2 corrected form · 8 neither30 transcription formfollowed the transcription
gemini10 corrected form · 10 neither20 corrected form · 10 neither27 transcription form · 3 neitherfollowed the transcription
gemini3120 corrected form · 10 neither20 corrected form · 10 neither30 transcription formfollowed the transcription
all models34 corrected · 68 neitherof 10244 corrected · 68 neitherof 112141 transcription · 9 neitherof 150—
Place of originp36 · place_of_origin
legnano for legnago		claude10 corrected form10 corrected form10 transcription formfollowed the transcription
claude_thinking9 corrected form · 1 neither8 corrected form · 1 neither10 transcription formfollowed the transcription
gpt7 corrected form—10 transcription formfollowed the transcription
gemini1 corrected form—10 transcription formfollowed the transcription
gemini31—10 corrected form10 transcription formfollowed the transcription
all models27 corrected · 1 neitherof 2828 corrected · 1 neitherof 2950 transcriptionof 50—
Surnamep28 · name
tedesco for todesco		claude3 corrected form · 5 neither10 corrected form——
claude_thinking1 corrected form · 1 transcription form · 6 neither1 transcription form · 7 neither8 transcription formfollowed the transcription
gpt——10 corrected form—
gemini7 transcription form5 transcription form10 transcription form—
gemini3110 transcription form10 transcription form10 transcription form—
all models4 corrected · 18 transcription · 11 neitherof 3310 corrected · 16 transcription · 7 neitherof 3310 corrected · 28 transcriptionof 38—
Given namep28 · name
argiè for argià		claude4 neither10 neither10 transcription form—
claude_thinking—1 neither10 transcription form—
gpt——10 transcription form—
gemini7 neither5 neither10 transcription form—
gemini3110 neither10 neither10 transcription form—
all models21 neitherof 2126 neitherof 2650 transcriptionof 50—
Arrival datep28 · arrival_date
february 9 for february 10		claude30 neither6 corrected form · 4 transcription form · 20 neither30 transcription formfollowed the transcription
claude_thinking28 neither20 transcription form · 7 neither30 transcription form—
gpt12 neither9 corrected form · 1 neither30 transcription formfollowed the transcription
gemini26 neither5 corrected form · 12 neither30 transcription formfollowed the transcription
gemini3130 neither18 corrected form · 12 neither30 transcription formfollowed the transcription
all models126 neitherof 12638 corrected · 24 transcription · 52 neitherof 114150 transcriptionof 150—

The transcriptions as collected

Emitted verbatim from transcriptions_as_collected/. Whitespace and tabs are preserved, including the layout runs that the corrected version replaces.

Page 28 — 20 February 1768 M.V. — 1769. 60 lines, verbatim as sent to the models.

20 Feb(brari)o 1768  M(ore) V(eneto) [February 1769]
[Left column]
Forastieri Ebrei annotati riferiti dal dep(utat)o
nella scad(ut)a Sett(iman)a come siegue  ~~

9 Feb(brari)o
Mandolin Todros di Padova in casa di
Salomon Levi dal Banco
		//-d(et)to
Leon q(uonda)m Iacob Tedesco con moglie, e cinque Fi
gliuoli di Francford nell’Osp(izi)o de’ Poveri
		//-d(et)to
Elia Amnon di Livorno: in Casa del q(uonda)m Be
niamin Ottolengo
		//-d(et)to
Salvador Anau di Ferrara con un Domes(tic)o
in casa del q(uonda)m Emanuel Pesaro
		//-d(et)to
Abram Asdà di Cento in Casa del q(uonda)m Simon
Todesco
			//13 d(et)to
Manassc Pincherle di Verona
Iosef Simon Volemburg d’Ocheners ambe
due in Casa del q(uonda)m Emanuel Pesaro
			//15 d(et)to
Israel, e David F(rate)lli Bassan di Verona: in
Casa d’Isach Morpurgo
			//16 d(et)to
Iseppo Grego con moglie una figliola ed
Un Domestico di Verona: in Casa di Benetto
Polacco

[Right column]
Forastieri Ebrei partiti e riferiti dal Dep(utat)o nella
Scad(ut)a Sett(iman)a come siegue ~ ~ ~
		12 Feb(brai)o
Aron Arcangiolo di Pesaro partito disse li 11
Scad(u)to allogiò dal q(uonda)m Beniamin Ottolengo
		// --d(et)to
Argiè q(uonda)m Leon Todesco con moglie di Francford 
p(ar)titi d(isse) li 11 scad(u)to allog(iaro)no all’Osp(izi)o de’Poveri 
			//13 d(et)to
David Levi Minzi di Montagnana
Leon Grassetti di Verona ambedue partiti disse
li 12 scad(u)to allog(iaro)no q(uonda)m Emanuel Pesaro
// --d(et)to
Isach Luzzato di S. [San] Vido partito disse li 12
scad(u)to allogiò da Samuel Capriles
// --d(et)to
Mandolin Todros
Regina Todros [this line and the above in curly bracket] di Padova
Partiti disse li 12 scad(u)to allog(iaro)no in
in Casa di Salomon Levi dal Banco
		// 14 --d(et)to
Salamon Cormos da Cormos partito disse li
13 scad(u)to allogiò in Casa del q(uonda)m Emanuel 
Pesaro
		--d(et)to
La Vedova Gentila di Fer(rar)a partita disse li
12 scad(u)to allogiò all’Osp(izi)o de’Poveri