The Semantic Gazetteer of the 1st Edition OS Maps

2 · The reconstructed alphabets

The Ordnance Survey's lettering, rebuilt

Each face is a new vector outline, reconstructed glyph by glyph from the OS's 1934 alphabet plates. The three antiquity hands lead; their assignments follow the 1881 rule, not the 1934 plates the letterforms were drawn from.

The shapes are dated too. Comparing these faces against the engraved sheets shows the letterforms drifted as well as the assignments. Old English survives the half century largely intact — plate and sheet agree. The face the OS came to call Lutheran did not: the 1934 plate prints a light, open, pointed letter, while the c.1888 sheets carry a denser, more angular hand for the same class. What follows is therefore a faithful record of the 1934 plates, not a facsimile of 1888 engraving, and the German Text specimen should be read with that gap in mind.

Antiquity assignments follow the verified 1881 six-inch column; the display-face uses below are summarised from the wider specimen record. The letterform comparison above is a visual reading of a small number of specimens; it is offered as an observation, not a measured result.

Type a Dartmoor name — the antiquity hands carry the thesis.

Reconstructed from the OS's 1934 alphabet plates (National Library of Scotland). New vector outlines; the 1934 letterforms are out of copyright.

3 · From detections to a name

The recognition pipeline, one label at a time

Recogniser output is fragmentary — unlinked tokens, mixed with survey furniture, and misread where the blackletter forms are unfamiliar. One real Dartmoor label, carried through each stage. Counts are for this label; the crop is unaltered.

4 · The documentary record

Which edition's rules apply

The period code was revised across editions, and the face names drifted while the periods were reassigned beneath them. Reading an 1888 sheet means fixing the rule to the instrument in force — drawn here from the primary documents held for this study.

5 · Measured results and future work

Where the numbers stand

Each figure is measured and carries its interval. A number that moved because the reference data was wrong is kept distinct from one that moved because the method was.

87.9%

merger precision · 95% CI 84.0–91.8 · order correct on 98.8% of true merges

Geometric name-merger, population-weighted

The raw adjudicated gold set scores 81.6%, but it was built to a quota heavy in four-word chains (15% against a true 6.5%), and precision falls with chain length. Weighting each band by its real share of the compounds gives 87.9% — the unweighted figure would misdirect the fine-tuning decision.

The "lost-names" set, checked by hand

84,305 records flagged as lost — the three denominators do not compose

A stratified 297-crop pixel check on what the pipeline flagged as lost. Each rate is measured against a different denominator, so all three are shown rather than multiplied. Checked against OS Open Names, the national heritage POI set and the statutory designation records; a name may survive in sources these do not cover.

Not a name — map furniture38.3%

Misread (of those that are names)26.9%

Genuine lost name45.1%

38,042 genuine lost names (95% CI 33,416–42,424); the national extrapolation falls from ~1M to ~450k. The earlier figure counted records the pipeline called lost, never checked against the pixels.

1 / 1,094

false removals in this window · 0.09% · corroborated against the c.1900 second-edition transcription

What the cleaning stage throws away

Half of all detections (1,094 of 2,094) are removed here. Audited exhaustively against the independent c.1900 transcription of the second edition: only 10 removed records contain four or more letters at all, and exactly one is corroborated as a real name — Tumulu#, a prehistoric burial mound the recogniser mangled. The single false removal is an antiquity — the one class this work exists to find. Losses are rare but not randomly distributed.

31

defective joins visible in 1,000 kept records (~3%)

How well the letters join

15 over-merges — survey furniture welded into pseudo-names (B. M. Guide Post), which then escape the abbreviation filter because the weld makes them too long. 16 confirmed under-merges, where both halves of a name survive apart (Smallcombe + Wood, 24 m). Both failure modes are structural and known, not random. Population precision remains the adjudicated 87.9%.

86,166

disagreements · 18.2% of kept records · shadow census, nothing published

The rulebook, applied in shadow

Every kept record was re-read against the OS's own written conventions, and the conventions disagreed with the string pipeline 86,166 times. The largest single class: 31,589 bench marks arriving as two records — invisible to a string rule, recoverable structurally. 57,793 records annotated; zero removal decisions changed. The census runs in shadow: no figure here has entered a published result, and no rule graduates without a measured before/after.

Corrections

Where a published figure was wrong, the corrected value replaces it everywhere and the superseded number is kept, with its cause, in a public log. Mistakes are published, not overwritten.

What this unlocks

Reading the survey's typography turns a picture of Britain into a queryable record of what its surveyors judged to be there.

i

Queryable Victorian Britain

Class search over the survey's own testimony — every workhouse, every mineral railway, every prehistoric antiquity. The engraved size of a name is the Survey's own weighting of its importance, and it is recoverable.

ii

A census of destruction

Surveyor-attested losses — the “Site of…” annotations — together with the genuine lost-name residue as dated leads, and the first-to-second-edition comparison that the matching layer now makes possible.

iii

Evidence-fused candidate sites

Name testimony, period-specific terrain expectation and statutory-record silence, ranked and referred privately to county archaeologists — never published as discoveries.

iv

Method that transfers

The conventions file and the edition-dating approach are not Dartmoor-specific: they apply to the 25-inch, to the Irish survey, and to any governed document series.

Future work

A typeface classifier

A typeface classifier is the next stage: a labelled training set drawn from the sheets themselves is in preparation, after which the readings above move from hand-verified examples to measured output. No automated classification is claimed on this page.

A national semantic gazetteer

The aim is to read every antiquity name on the first-edition six-inch through the OS's own period code, dated to the edition in force, in collaboration with the National Library of Scotland.

The papers

A methods paper on typeface as testimony is planned, with the reconstructed OS alphabets released alongside it for reuse.

The code, in the field