What the extraction is worth

Generated by bourbaki report extraction -write. Do not edit.

Every number here is counted off the page files in pages/ and the volumes in manifests/books.yaml, so it is what the corpus holds today and not what anybody hopes it holds. A volume with no row in the coverage table has no page file at all.

Read is how many pages of the PDF have a page file, blanks included, since a blank page that has been looked at is a page that is done. Checked is the pages the validation rules ran over, which is the read pages that are not blank, and accepted is how many of those pass every rule that ran. The rules are the same ones the extraction itself accepts a page on, so this is that decision re-run against what is on disk.

Passing the rules is not the same as being right. A page can balance its dollars, carry a plausible running head and still read an interval as a set, which is what ocr audit is for. A rejected page is a place to look and most of them are one character.

The acceptance column is not quite reproducible and it is worth knowing why before the numbers are quoted. A render manifest under images/ is what says a page is blank or nearly blank, and the short rule is relaxed on a page it calls sparse. images/ is not in git, so a clean checkout has no manifest for any volume and counts a handful of sparse pages as too short where the machine that did the extraction does not. The run that wrote this file had a manifest for 20 of the 20 volumes that have pages on disk. Render the PDFs and the figure goes up by a tenth of a point or so; it does not move any other column.

Coverage, worst first

Volume	Lang	Text layer	PDF pages	Read	Unread	Coverage	Checked	Accepted
ac-i-iv-fr	fr	ocr	361	0	361	0.0 %	0	none checked
ac-i-vii	en	none	642	0	642	0.0 %	0	none checked
ac-v-vii-fr	fr	ocr	346	0	346	0.0 %	0	none checked
ac-viii-ix-fr	fr	ocr	204	0	204	0.0 %	0	none checked
ac-x-fr	fr	ocr	180	0	180	0.0 %	0	none checked
alg-i-iii-fr	fr	ocr	645	0	645	0.0 %	0	none checked
alg-iv-vii-fr	fr	ocr	426	0	426	0.0 %	0	none checked
alg-ix-fr	fr	ocr	211	0	211	0.0 %	0	none checked
evt-i-v-fr	fr	ocr	372	0	372	0.0 %	0	none checked
fvr-i-vii-fr	fr	ocr	329	0	329	0.0 %	0	none checked
hist	en	ocr	299	0	299	0.0 %	0	none checked
int-i-iv-fr	fr	ocr	293	0	293	0.0 %	0	none checked
int-ix-fr	fr	ocr	131	0	131	0.0 %	0	none checked
int-v-fr	fr	ocr	159	0	159	0.0 %	0	none checked
lie-i-fr	fr	ocr	144	0	144	0.0 %	0	none checked
lie-ii-iii-fr	fr	ocr	318	0	318	0.0 %	0	none checked
lie-iv-vi-fr	fr	ocr	284	0	284	0.0 %	0	none checked
lie-vii-viii-fr	fr	ocr	61	0	61	0.0 %	0	none checked
top-i-iv	en	ocr	443	0	443	0.0 %	0	none checked
top-i-iv-fr	fr	ocr	367	0	367	0.0 %	0	none checked
top-v-x	en	none	372	0	372	0.0 %	0	none checked
top-v-x-fr	fr	ocr	329	0	329	0.0 %	0	none checked
var-fr	fr	ocr	190	0	190	0.0 %	0	none checked
int-vii-ix	en	ocr	332	1	331	0.3 %	0	none checked
alg-iv-vii	en	ocr	460	2	458	0.4 %	0	none checked
int-vii-viii-fr	fr	ocr	224	1	223	0.4 %	0	none checked
hist-fr	fr	ocr	374	2	372	0.5 %	1	100.0 %
evt-i-v	en	ocr	368	2	366	0.5 %	1	100.0 %
int-i-vi	en	ocr	487	3	484	0.6 %	2	50.0 %
lie-ix-fr	fr	ocr	141	1	140	0.7 %	1	100.0 %
alg-i-iii	en	ocr	734	8	726	1.1 %	3	100.0 %
int-vi-fr	fr	ocr	106	2	104	1.9 %	1	100.0 %
lie-i-iii	en	ocr	481	15	466	3.1 %	5	100.0 %
lie-iv-vi	en	ocr	314	13	301	4.1 %	0	none checked
fvr-i-vii	en	ocr	354	15	339	4.2 %	5	80.0 %
alg-x-fr	fr	none	222	57	165	25.7 %	57	100.0 %
ens-i-iv	en	ocr	418	408	10	97.6 %	407	100.0 %
alg-viii	en	native	505	505	0	100.0 %	494	99.4 %
alg-viii-fr	fr	native	487	487	0	100.0 %	486	99.2 %
lie-vii-ix	en	native	439	439	0	100.0 %	439	97.7 %
ta-i-iv-fr	fr	native	512	512	0	100.0 %	510	99.2 %
ts-i-ii-fr	fr	native	346	346	0	100.0 %	344	99.1 %
ts-iii-v-fr	fr	native	579	579	0	100.0 %	576	99.5 %
all			14989	3398	11591	22.7 %	3332	99.1 %

Of 43 volumes, 6 are read through, 14 are begun and 23 have no page file at all. That is worth saying next to the percentage, because 22.7 % of the library read could mean every volume a fifth of the way through and it does not: it means a few volumes finished and most of them never opened.

How the pages were read

Volume	native	ocr	blank	flagged	by hand
int-vii-ix	0	0	1	0	0
alg-iv-vii	0	0	2	0	0
int-vii-viii-fr	0	0	1	0	0
hist-fr	0	1	1	0	0
evt-i-v	0	1	1	0	0
int-i-vi	0	2	1	0	0
lie-ix-fr	0	1	0	0	0
alg-i-iii	0	3	5	1	0
int-vi-fr	0	1	1	0	0
lie-i-iii	0	5	10	2	0
lie-iv-vi	0	0	13	0	0
fvr-i-vii	0	5	10	0	0
alg-x-fr	0	57	0	3	0
ens-i-iv	0	407	1	86	52
alg-viii	494	0	11	52	40
alg-viii-fr	486	0	1	32	14
lie-vii-ix	427	12	0	17	7
ta-i-iv-fr	510	0	2	28	4
ts-i-ii-fr	344	0	2	16	6
ts-iii-v-fr	576	0	3	18	1
all	2837	495	66	255	124

Volumes with no page file are left out of this table and are in the one above, where the zero is the point. Native is the text layer of a born-digital PDF, ocr is a model reading a page image, ocr-repair is a model putting back a display the text layer lost, and blank is a page with nothing on it. Flagged is the pages the reading was not sure of and by hand is the pages somebody has since repaired themselves, which extraction is forbidden to write over.

Pages left at ocr-failed

None. The extraction rejects a page it cannot read rather than committing it, so a page that fails three times leaves no file behind and shows up above as an unread page. S06 in the audit is what would catch one if a future run wrote it.

What the rules reject

Rule	Pages
short	19
head	9
leak	1

One page is counted once per rule it fails, so these add up to more than the rejected pages when a page fails two.

Volume	Rejected	Rules	Pages
int-i-vi	1	head	54
fvr-i-vii	1	leak	2
alg-viii	3	short	1 2 3
alg-viii-fr	4	short	2 3 4 9
lie-vii-ix	10	head, short	1 2 147 216 229 270 301 310 312 319
ta-i-iv-fr	4	short	1 2 4 355
ts-i-ii-fr	3	short	1 2 3
ts-iii-v-fr	3	short	1 2 3

Where two of the rules did not run

There is no page map for alg-x-fr, so the running head rule and the page label rule were skipped over those pages. Six of the eight rules ran, and the acceptance figure there is worth that much less than the rest of the column.