Knuth-Plass Line Breaking With River Penalties

Baseline paragraph breaking uses classic box/glue/penalty scoring. Extension adds vertical space-alignment penalties so justified lines can avoid visible rivers while still preserving Knuth-Plass style badness and fitness evaluation.

explicit spacing per line post-hoc detected river overlay beam search keeps search tractable

Paragraph Input

Display Mode

Baseline only River-prevention only Side-by-side comparison

Controls

Paragraph width42 em

River penalty weight240

Alignment tolerance0.55 em

River zone0.75 em

Lookback depth k2

Beam width18

Min chain length3

Min river score1.4

Max drift / line1.25 em

Toggles

River highlighting Neural (live) Demerit breakdown display Debug river geometry Raster refinement

Neural Detector Overlay

Precomputed predictions from the focal-loss U-Net on hand-labeled test paragraphs. The threshold slider binarizes the saved probability heatmap client-side.

Sample

Layers

Neural heatmap Neural binary Heuristic mask Ground truth

Threshold0.50

Overlay opacity0.75

neural (heatmap / binary) heuristic mask ground truth

No overlay assets found. Run python3 scripts/export_web_samples.py from the repo root.

Detector-Guided Reranking

Generate a small set of candidate layouts with different river-penalty settings, score each on classical demerits + river severity, and pick the best. River severity is a geometry surrogate — it uses the same chain features the neural detector was trained on, but does not run the U-Net live. See docs/detector-guided-reranking.md.

River weight α1.00