In 1803 the bakers of Dundee sued their own town council over a mill monopoly. For two centuries their Information — twenty pages of dense letterpress, argued by George Joseph Bell — sat in a bound volume, findable only by someone who already knew to look for it. It is findable now: by the bakers' name, by the law of thirlage, by the mills of Dundee on a map, by every person and place its pages mention. This essay is about why that took new methods — and why we believe those methods matter beyond this archive.
The arithmetic of description
The Session Papers resist the usual remedies. The University of Virginia Law Library holds fifty-eight linear feet of them; with the digitized holdings of the University of Edinburgh and the Library of Congress, the unified corpus runs to tens of thousands of printed documents. Since 2018, this project's ambition has been to describe them at the level of people, places, and subjects — not just case names — because that is where the social history lives: the merchants and midwives in the testimony, the farms in the rent rolls, the sugar and linen in the disputes.
Our catalogers spent roughly eight years doing exactly that, superbly, and described about 4,500 documents. That work is the foundation of everything here. It is also arithmetic against a wall. Close description takes minutes to hours per document. Reading the bodies of the documents multiplies the work: we measure about fifty entity mentions per document once page-level reading is in play — roughly 1.75 million mentions corpus-wide. If a person validated each uncertain one at thirty seconds apiece, that alone is more than six person-years of clicking that produces no new knowledge. No archive gets that labor. The honest choices were three: describe a fraction and stop; describe everything shallowly; or build something new.
What we actually built
The design brief was never "replace the catalogers." It was almost the opposite: convert what the catalogers had already built — the vocabularies, the authority records, the house style, thousands of confirmed case links — into calibration data and matching authorities for machines; run the machines autonomously wherever their measured precision earns it; and route to human judgment only the decisions that actually create knowledge. Machine proposals are scored against the catalogers' own records before any pipeline is trusted; every assertion carries its evidence and a confidence; catalogers confirm, correct, or reject from queues built for the purpose. The machine supplies reach. People supply judgment. The decade of expert work is not superseded by the computation — it is the thing that makes the computation trustworthy.
Two further commitments shape the system. Everything in production runs on open-weight, open-source models on university infrastructure: the methods can be inspected, the results reproduced, and the documents never leave our custody — an archival argument as much as a technical one. And everything the machine asserts is provenance-stamped and reversible, layered above preserved originals it can never alter. A future scholar can disagree with every one of our inferences and still have the full digitized record, untouched, to build on.
Why it matters beyond Scotland
The Session Papers are a spectacular corpus — the eighteenth-century British Atlantic arguing with itself, one dispute at a time — and unifying the scattered holdings into a single searchable collection is reason enough for the work. But the predicament this project answers is not Scottish. Every large archive faces the same arithmetic: expert description does not scale, and the sources most valuable for social history — dense, messy, name-rich — are precisely the ones that need description most. What this project demonstrates is a way through that respects the profession: expert knowledge as the calibration standard, open models as the workforce, provenance as the contract, and human judgment kept exactly where it creates knowledge. The bakers of Dundee are findable now. So is the method.
