Relaxation Labeling Reads the Biblea hierarchical-relaxation-labeling NLP engine, set upon the King James scripture

The same engine that tracks motion-capture markers and grades physics claims here judges pairs of verses by natural-language inference — agree = P(entail) − P(contradict) — to surface contradictions, kindred passages, and themes. Scholarly textual analysis; no verdict on faith.

▶ Watch it settle ◆ 2D convergence ▦ Compatibility tensor ⊙ Signal space ⟋ Convergence curves ✦ Verse explorer

The Contradiction Ladder — 29 famous pairs, ranked by the machine

contradiction (negative) harmony (positive) · ◆ blind = a real contradiction the machine failed to see · click a row to read the verses

Where the Machine Goes Blind — real contradictions it rated harmless (or harmonious)

The judge detects the grammar of contradiction (negations, swapped tokens), but is fooled by surface-rhyme and blind to narrative discrepancy.

The Web of Scripture — top contradictions & concords across all 66 books

Each node a verse, ordered Genesis → Revelation and coloured by section of scripture. Red arcs join contradictions; green arcs join kindred verses. Tensions span the whole canon — Law to Apocalypse.

Themes — 14 clusters the engine found by relaxation over verse vocabulary

Method. KJV (Project Gutenberg #10) parsed to 31,102 verses. Candidate pairs blocked by TF-IDF nearest-neighbors, judged by a cached DeBERTa-MNLI model (hrl.nli), segmented by KMeans, and (for fault-lines) relaxed by hrl.consensus. Fully offline. Built by the Gospel-HRL pipeline.

“The machine sees the grammar of contradiction, but not the facts.”