Afterlives: quotation in Classica Codex
High confidence means the match passed rules that held up when samples were read one at a time. Two AI models each read a random sample of 40 plain citations. Claude accepted 39 as quotations or close verse echoes, though a stricter reader might call some of those stock phrases. Codex, from OpenAI, read a different 40 and accepted 35. Of its other five, two were real quotations credited to the wrong source, one was a coincidence, one was an editor's note and one it could not settle. So expect about one row in ten to be wrong. Low-confidence leads are weaker: in Claude's samples roughly a third are real, so read them before relying on them.
A lot is missing. On a held-out part of Loci Similes, a benchmark compiled by experts, the matcher found about 45% of direct citations counting the low-confidence leads, somewhat fewer without them, and almost no allusions. A line with nothing here may still have been quoted. Only the Patrologia Latina and CSEL were searched, against classical authors who died by 300 CE. Classical writers quoting each other are not included.
Writer counts are upper limits. Some texts sit under two names, and some writers copy others. A passage with the same wording around the quotation as an earlier witness is marked and left out of the count, but that test will not catch every case. Commentators such as Boethius on the Topica quote because that is their job, and they are not set apart yet.
The Latin is as scanned, so Migne's misprints stay in. A few works were written in Greek, so their Latin is a later translation. Constantine's Oration does quote Virgil, for example, but its wording here comes from a later edition. Some notes by early-modern editors remain under ancient names, beyond the ones flagged. Authors' names use common English forms where they exist, and a work may not really be by the author the library names. Dates are conventional and approximate, and they belong to the named author. References to the later writers are the library's own numbers, which may not match a printed edition. Links for the Patrologia and CSEL open the whole file, not the passage.
For one line followed in depth, with English, see Afterlives of a Line.
This is AI work, and no specialist has reviewed it. Claude, an AI model from Anthropic, wrote the matcher, set its rules, read the samples and wrote these pages. Where a page says something was read, Claude read it. A person chose the project and decided to publish it.
The matcher compared about 1.65 million sentences from the Patrologia Latina and CSEL with classical Latin by two routes. One needs several rare words close together in both texts. The other looks for short runs of consecutive words that are rare in both the classics and Christian Latin. Each row passed one of those rules; no person or AI has read most of them.
Scripts took every excerpt and source line directly from the library's text, so the Latin here is not retyped. Scripts also checked the Scaife references against the TEI files, and a sample of the links was tried in a browser. Most links have not been opened.
Juvenal's Latin is not shown. The Perseus file used here has the placeholder “Test” where its licence should be, so his lines appear as references with a link to read them at Scaife. Passages that quote him are shown as usual.
The benchmark is Loci Similes (Apache-2.0). It measured the matcher, and none of its text appears here.
The text and data of this page are shared under CC BY-SA 4.0, as its sources require. The texts were read in Classica Codex.