cognator

Framework-native cognate detection with calibrated confidence

cognator clusters phonologically related forms across lects within a concept, emitting per-prediction confidence backed by a held-out calibration. The full pipeline — merkmal-based phonetic prior, asymmetric alignment, support-aware graph construction, Infomap clustering, and form-local risk bands — runs in your browser as a single WebAssembly module. No installation, no server-side processing.

Framework-native scoringmerkmal phonological features, asymmetric lect-pair scorers, prosody-aware alignment

Calibrated risk bandsEvery prediction labelled low / medium / high risk against a transferable Dev-Core fit

Fast29–54× faster than LingPy on CoreCog; 18 MB WASM, 3 MB gzipped

OpenC core with a Python package and C/WASM browser build; same pipeline as the CLI

Run a detection

Loading cognator engine…

Advanced configuration

Preset

Seed

Scoring

Scorer

Iterations (EM)

Lambda (learned)

α₀ concentration

k sharpness

α₀ scaling

Alignment

Class-reduce (NFD strip)

Null

Shuffle runs

Min pairs (learned)

Threshold

Threshold (0 = Otsu)

Mode

Min clamp

Max clamp

Shrink

Graph construction

Graph mode

Expansion

Top-k

Adaptive margin

Adaptive either-max

Adaptive ambiguity

Mutual boost

Triangle boost

Contrastive penalty

Contrastive floor

Risk weight: rank

Risk weight: gap

Unilateral penalty

Rival floor

Refinement

Mode

Confidence threshold

Bridge ratio

Bootstrap & clustering scope

Bootstrap runs (0=off)

Bootstrap σ

Scope

Rows colour-coded by calibrated risk band (Dev-Core fit). Click column headers to sort.

Concept: Nodes coloured by predicted cluster · border by risk band

low risk medium risk high risk uncalibrated

cognator

Run a detection

Guide Close

Guide