Method and limits — From Privilege to Patent

What was counted

English invention privileges and patents come from two lists: Hulme (1896, 1900), who drew 55 industrial monopoly grants of 1561–1603 from the calendars of the Patent Rolls, and Woodcroft (1854), whose chronological list numbers 1 to 14,359 for 1617–1852. 1604–1616 falls between the two lists. Hulme (1896) notes that a return of grants made before 1623, ordered by the Commons in 1827, was never presented.

The Stationers' Register is read straight from the Stationers' Register Online TEI (1557–1640, 10,453 entries). Venetian printing privileges are Fulin's 256 documents (1882), read and coded one by one. For 14 of those coded as privileges, Fulin's summary records the petition but not the grant; they were counted on the strength of the penalty clause or of editions 'con privilegio' cited in Fulin's notes, and the note column says so. The Dutch figures are a recount of the data published by Buning and Looijesteijn.

Venetian invention patents were not counted here. The record-level source checked here is Berveglieri's printed calendar (2020); earlier, Mandich (1936, 1948) transcribed and classified 109 grants of 1474–1550 (Comino et al. 2017). Figures from the literature are marked 'secondary' and kept apart.

Reading Woodcroft's list

The word coordinates of the Internet Archive OCR were used to separate the margin column (surnames, sealing dates) from the text column (numbers and abstracts). Because numbers must run from 1 to 14,359, the longest rising chain of number-like lines was kept and candidates caught in its gaps were renumbered in order. 115 pages on which the OCR lost entry numbers were read again from the page images with tesseract 5. 36 numbers still have no text: they are counted in the yearly totals (with their neighbors' year) and left out of metrics that need the text, except the share of extension entries, which keeps them in the denominator as entries that are not extensions and is therefore a lower bound.

Dates are the sealing dates printed in the margin. For 13,611 entries the margin date is used as read. Margin dates damaged by the OCR (351 entries) are used only when their year falls between the clean dates before and after, and the day is kept only when its digits are clean. For 15 entries whose margin date the OCR lost altogether, the date was read on the page image (data/coding/woodcroft_date_overrides.csv). Otherwise the neighbors' year was used when both neighbors agreed (382 entries).

A year that differs from both neighbors while they agree with each other is treated as an OCR misreading. Nos. 7046 and 7049 are printed '29th March 1835' in Woodcroft itself among entries of March 1836; they are taken as misprints and counted in 1836.

A margin line containing digits is a date, not a grantee's surname: the number of grantees (shared grants) counts only margin lines without digits. An address abroad is a place outside the British Isles in the grantee's description; London street names such as 'Hanover Square' and memberships of foreign societies do not count.

The term is the first term stated in the abstract, read in early-modern spellings (fourteene, one and twentie), Roman numerals (xiiij, xxjtie) and OCR-damaged forms (xilij). Where an entry states further terms (mines, Ireland, a later revenue), they are kept in the column term_other (4 entries in 1617–1700). For nos. 1–126 (1617–1640) the first term agrees in every case with an independent reading of the page images. From about 1820 the abstracts stop stating the term, so later terms are left blank; the statutory 14 years was not assumed.

Rents, provisos and foreign links are judged from the words of the abstract. The rent metric is the share of entries with rent wording, not a legal finding that rent was owed.

Field classification

There are 17 fields. Word rules are applied to the title of the invention in a fixed order and the first match is taken. Entries before 1700 do not set the title off in quotation marks, so the whole abstract is used, after early-modern spellings are normalized.

The rules were fixed before any results were seen but were revised three times before publication. r2 removed the bare word 'instrument' (any device, in 17th-century usage) from the instruments field; r3 added the spelling table and widened the text used for entries before 1700. These two followed an implausible first result.

Checked against an independent coder's recoding of 212 entries from the text (R2), the r3 rules agreed on 144/212 (68%, 95% CI 61–74%) and on only 2 of 11 mining entries: 'tan pits', 'pit coal' and 'smelting copper ore' had gone to mining and gas lighting to chemicals. After the r4 fix, precision was measured again on a fresh sample of 212 entries drawn with another seed (R2b), not on the sample that prompted the fix. The field figures on the reliability cards are from R2b. The revisions are recorded at the head of src/fields.py.

Reliability cards

Image check (R1): an independent coder read 120 entries, 30 per period, from the page images (number, date, number of grantees, 'communication', term) for comparison with the extracted values. Year agreement 119/120 (99%, 95% CI 95–100%), full date 112/120 (93%, 95% CI 87–97%), term 118/120 (98%, 95% CI 94–100%), number of grantees 118/120 (98%, 95% CI 94–100%), communication 119/120 (99%, 95% CI 95–100%).

Text recoding (R2b): another coder coded 212 entries (11 per field, 25 'other') from the text alone: field, address abroad, communication, term. Field agreement 145/212 (68%, 95% CI 62–74%) (κ 0.67; the sample takes equal numbers per field, so per-field precision weighted by actual field shares gives 67%), address abroad 206/212 (97%, 95% CI 94–99%), communication 208/212 (98%, 95% CI 95–99%), term 210/212 (99%, 95% CI 97–100%).

Fulin recoding (R3): a second coder recoded 40 random documents under the same codebook. Agreement on privilege or not 40/40 (100%, 95% CI 91–100%), term 40/40 (100%, 95% CI 91–100%), beneficiary type 37/40 (92%, 95% CI 80–97%).

Hulme's 55 grants: another coder read the page images independently; dates agree 55/55 (100%, 95% CI 93–100%) and terms 55/55 (100%, 95% CI 93–100%). The decade counts match Hulme's own table on p. 52 (23, 12, 14, 6, 0).

Checks against prior figures

Woodcroft's numbered entries come to 2,239 for 1660–1799 against MacLeod's (1988) 2,240 (a difference of 1), and 13,070 for 1700–1850 against the 13,070 of Nuvolari et al. (2021) (a difference of 0). The counting rules of those studies (whether extensions are included, for instance) were not checked, so the source of the differences is not settled.

The decade totals of the Stationers' Register match the counts returned by the SRO website's search API (2026-11-03; data/sro/sro_site_decade_counts.csv) in every decade. The Fulin count of 240 privileges for 1492–1526 is 14 below Kostylo's (2010) 254, possibly because of what each counts as a privilege.

What this site does not do

It makes no causal estimates. Years such as 1474, 1624 and 1710 are reference points for before-and-after comparison only; the site does not say that a law increased patenting.

Record-level counts of Venetian invention patents, French royal privileges, the Stationers' Register after 1640 and United States federal patents (from 1790) are candidates for later versions.

Versions

1.0 (2026-11-03) first release.

1.0.1 (2026-11-03) corrections after independent verification. The Dutch foreign-origin codes are read correctly (8 and 9 mark Dutch grantees). Woodcroft's 17th-century terms (old spellings, Roman numerals) and damaged margin dates are read better, leaving no entry without a year. Margin date lines are no longer counted as grantees, and London street names are no longer read as addresses abroad. The 'last 21-year term' statement is replaced by a table for 1617–1700. The reliability cards were measured again on the corrected values.

1.0.2 (2026-11-04) Four statements in the text now match the data: the denominator of the extension share (36 numbers lost their text, and they stay in the denominator), the range of Elizabethan terms (6–30 years), the range of the shared-grant share from the 18th century (3.6–17.1%) and the name of the term column. The language switch leads to the same page, and table headers and chart labels are given in each language. Shares are shown as percentages, and long values wrap on narrow screens. The steps to rebuild from the public repository were corrected. No metric values changed.