Isaacus open-sources a legal conflict detector: encoder-only pipeline reviews contracts in under 20s

Neon0asis · reddit · 2026-11-16

Isaacus open-sourced (MIT) a legal document conflict detector that uses small discriminative encoder models instead of generative LLMs for clause-level pairwise review.

Why not LLMs: pairwise clause comparison is quadratic — two 133-clause contracts mean 17,689 API calls; dumping files into Claude/GPT yields weak recall and hallucinations.

Pipeline: (1) parse documents into hierarchical clause trees, (2) embed clauses, (3) zero-shot legal classifier drops non-provisions, (4) shortlist top-3 matches with cosine similarity ≥0.5, (5) reranker scores relevance + NLI classifier scores entailment, (6) same-topic high conflict = deviation, no match = gap.

Longest document pairs ran in under 20 seconds. An optional LLM step only summarizes/redrafts findings, so it can't hallucinate conflicts into existence. Demoed on Bob Iger's 2011 vs 2022 Disney employment agreements, finding the 2011 version more favorable (e.g., no IP assignment clause).

Original post →

More from Apps

Apps channel →