regulae

Source

Systematic sound correspondences between lects, discovered from cognate data.

Statistics & options — extra checks, slower, off by default
change only for a controlled comparison

Early release. Tested on fifty-six corpora, including real Latin/Spanish and Proto-Polynesian/Hawaiian. Anomaly detection isn't ported yet. The resampling checks stay off by default because they retrain the model many times; see capabilities.

What this does

  • Aligns cognate forms across two or more lects
  • Finds which segments answer to which, and how often
  • Discovers the environments a correspondence is conditioned by
  • Ranks cognate sets by how badly they fit what it learned

What this does not do

  • Decide which words are cognate (that's your input)
  • Reconstruct proto-forms
  • Claim a direction of change: correspondences are written a ~ b, never a → b
  • Build family trees or date anything
1

Load cognates

One row per meaning, one column per lect. Whole words — segmentation is automatic.

2

Run

Get a table of correspondence classes: which segment answers which, and where.

3

Read the table

Select a correspondence to see the alignments it rests on. Open a tab for events, outliers, or cross-dimensional rules.

Correspondence classes

Select one to see the alignments it rests on.

correspondence count sets rate · 95% CI

Select a correspondence to see the cognate sets it rests on.

Classes that look like one change

eventcountsets

Residue

cognate setzcost/segmentconfidence

Cross-dimensional rules

rulecountvs.

Provenance, to cite and reproduce this run

A snapshot for inspection, not an interchange format. It holds one best answer, without the uncertainty a historical claim needs.