Analysis outputs and model variants
The current paper-facing analysis uses the active three-bucket tags, then restricts the model to mythic vs. historical sentences.
| Vocabulary | Book Scope | Rhetoric Markers | Sample | Features | Accuracy | Page |
|---|---|---|---|---|---|---|
| Lemma | Excluding 4 and 8 | Included | 5,424 | 1,000 | 0.692 | Open |
| Lemma | Excluding 4 and 8 | Removed | 5,424 | 1,000 | 0.693 | Open |
| Lemma | All books | Included | 7,032 | 1,000 | 0.728 | Open |
| Lemma | All books | Removed | 7,032 | 1,000 | 0.716 | Open |
| Surface | Excluding 4 and 8 | Included | 5,424 | 1,000 | 0.712 | Open |
| Surface | Excluding 4 and 8 | Removed | 5,424 | 1,000 | 0.698 | Open |
| Surface | All books | Included | 7,032 | 1,000 | 0.723 | Open |
| Surface | All books | Removed | 7,032 | 1,000 | 0.726 | Open |
480 sections and 1,734 numbered sentences processed; 3,474 mention rows extracted.
Open People ReportPassage-level morphosyntactic features from the gpt-5.4-mini grammar parser, with word-frequency and character n-gram baselines.
Open StylometryFeature-family projection display for comparing current parsed passages and mouseover neighbor context.
Open ProjectionCross-validated mythic, historical, and other sentence classifiers using grammar-derived stylometric features.
Open ClassifiersRidge regressors that test whether stylometric features can predict sentence book number better than a mean-book baseline.
Open RegressionBook-by-book proportions for the strongest morphosyntax regression features, plus sentence-length KDE and violin diagnostics.
Open Trends1,200 tagged parsed sentences across 6 discourse modes.
Open Mode ControlReruns the main lemma model after removing kinship, reporting, memorial, and military-political vocabularies.
Open Ablation AnalysisTrains on all but one book, then tests whether the model generalises to the held-out book.
Open Held-Out AnalysisConfident false mythic and false historical predictions for close-reading follow-up.
Open Error AnalysisConfusion matrix, label-scenario refits, and noise stress tests using the Greta/Rosie Book 3 manual labels.
Open Sensitivity AnalysisGreek predictors of length residuals, plus the dual view of English terms found in longer or shorter-than-expected passages.
Open Translation Length AnalysisExploratory comparison of length-residual terms against the main mythic/historical classifier coefficients.
Open Diagnostic PageGeographic map, place-pair distances, and proper-noun co-occurrence network pages.
Open Place and Network PagesMythic/historical subgraphs, bridge nouns, book-level drift, and place-person/deity networks.
Open Network Analyses