Isaacus releases open-source legal conflict detector using encoder models and NLI

Details

Isaacus released an open-source legal conflict detector that identifies deviations and gaps between document pairs using encoder models and natural language inference (NLI).
The pipeline processes the longest sampled document pairs in under 20 seconds, avoiding the quadratic cost of generative models which would require 17,689 API calls for two documents with 133 clauses each.
The system enriches documents into hierarchical trees, embeds clauses as vectors, and filters non-provisions using a legal zero-shot classifier.
It shortlists the three most relevant pairs with a cosine similarity of at least 0.5, then scores them using a reranker for relevance and a classifier for entailment.
An optional LLM step can summarize findings or redraft clauses, but only after the encoder stack has identified conflicts, preventing hallucinated issues.
A demo using Bob Iger's 2011 and 2022 Disney employment agreements revealed that the 2011 agreement was more favorable to Iger and lacked an intellectual property assignment clause present in the 2022 version.
The code is MIT-licensed and available on GitHub, allowing users to swap in their own models or modify the architecture.

Key point