Draft — figures and text are still being checked. Not indexed; please do not cite yet.
Footnotes to Plato

Methods

How the numbers were made, and how far they can be pushed

This page is the short version. The full pipeline and methods record documents the build as actually executed — including the bugs found, the assumptions tested, and the things still outstanding.

The pipeline in brief#

  1. Delivery. A 381 MB gzipped full-text corpus from JSTOR Text Analysis Support: 69,177 documents across ten journals.
  2. Extraction. Candidate citation strings are pulled from article body text by shape — Stephanus (509d), Bekker (1097b12), and named-work forms. Roughly 141,000 genuine candidates survive filtering.
  3. Resolution. Each candidate is matched against reference tables of work boundaries. Where the locus falls unambiguously inside one work, it is resolved. Where it does not, it goes to a review queue with the competing candidates recorded.
  4. Aggregation. Resolved citations are counted per work, per passage, per journal and per year, and joined to article metadata for DOI links.

The governing principle is precision first: at every stage the default is to decline to resolve rather than guess, and every exclusion diverts data to an auditable file rather than deleting it.

What "placed" means#

Of about 141,000 candidates, 90,502 (64%) were placed to a specific work and passage. The other 36% are not errors and are not lost — they are recorded in the review queue with their candidate works.

The dominant reason for non-placement is collision. Stephanus pagination runs continuously across the Platonic corpus, so a citation to "80b" with no work named could belong to Meno, Apology, or several others. Where the surrounding text does not disambiguate, the pipeline declines to choose.

This has an important consequence for reading the charts: placement rate is a property of a work's page range, not of its scholarly importance. Crito places at 20% not because it is neglected but because its Stephanus range is crowded.

The two tiers#

Works are split at an 80% placement rate:

The 80% cut was chosen from the actual distribution, which is genuinely bimodal — a dense mass below 50%, an empty valley between 50% and 79%, and a spike from 80% up. The threshold sits in the valley rather than at a round number picked in advance.

Uncertain-tier works are included, not hidden. Excluding them would remove Phaedo, Phaedrus, Gorgias, Apology, Meno and Crito — 28% of all placed citations, and much of what the field actually argues about. They are marked instead: dashed lines in the time chart, a fading tail in the summary chart, and a stated placement rate wherever they appear.

What is deliberately not shown#

Known limitations#

Checking the work#

The derived data is published under CC BY 4.0, and the pipeline is open source. If you find an error, please tell me — corrections will be recorded and credited.