Every document in this corpus of Ottoman court records (kadı sicilleri) has been scored against thirteen psychological measures. Each score is the cosine similarity between the document’s text and a questionnaire measuring that construct, produced by a language model fine-tuned on Ottoman Turkish. A higher score means the document’s language sits closer to how that construct is expressed. These are textual measurements, not measurements of what anyone in the record believed.
Raw cosine values are tiny (roughly −0.09 to +0.09) and cannot sensibly be compared across measures, so panel ii offers two rescalings. Z-score expresses each value as standard deviations from the corpus mean — 0 is the corpus average, +1 is one SD above it — and is the default. Percentile gives the document’s rank from 0 to 100, which is easier to read but hides how large the gaps are. Both are computed over all 54,670 documents and do not change when you filter.
Click a dot to add or remove that court from the selection; click again to undo. With nothing selected, all courts are included. Dot colour shows the confessional majority recorded for the court and dot size the number of documents. Clicking a court name — on the map, in the side list, or in the records table — opens its full metadata with sources. Seven administrative courts have no coordinates and appear in the list beside the map.
Set the filters to describe one group and press Capture current filters under A; change the filters and capture B. Pressing Compute reports, for each selected measure, Cohen’s d with a 95% confidence interval and Welch’s unequal-variance t-test, with p-values FDR-corrected across the measures shown. Positive d means the measure is higher in Group A. Conventional readings are 0.20 small, 0.50 medium, 0.80 large. Because the groups are large, p-values are near zero almost everywhere and tell you little; the effect size and its interval are what to report. The preset buttons rebuild the comparisons from the ma‘rûz case study.
Twelve legal milestones — from the Millet system through the 1917 Law of Family Rights — are drawn over the temporal chart as numbered markers, with reform periods shown as shaded bands. Because eight of them fall between 1839 and 1876, and two share the year 1858, the markers stack into rows so that none is hidden. They are context for reading a trend, not analysis: the chart makes no claim that a change in scores follows from a change in law. Milestones lying outside the range currently plotted are listed below the chart and marked off chart; the list updates as you filter.
Dates are volume-level. Only 6,829 documents carry a single year; the rest carry a span, and the time chart places those at the midpoint of the span. Confessional majority is a two-category attribution for a whole court, and the metadata itself flags several as not securely sourced — treat it as a rough stratification, not a measured ratio.
Every section header collapses to a single bar when clicked, and expands again on a second click — useful for putting the map or the records table out of the way while you work on something else. Collapse all and Expand all sit at the top right. The map has its own Taller button when you need more room to aim at overlapping courts. In the filter panel, the Apply button turns red when you have changed something that has not taken effect yet, and pressing Enter in any number field applies immediately.
Every chart has Save PNG; the records table and the comparison table both download as CSV, carrying whichever unit is currently displayed.
not set
not set