Knowledge Pull Requests for Continual Document Authoring

Knowledge Pull Requests, or KPRs, introduce a reviewable alternative to opaque document rewriting. The system decomposes source and main documents into atomic, decontextualized claims, classifies them as supported, absent, or conflicting, filters for coverage and relevance, routes new claims to sections, and rewrites only affected sections. Its ChangeLog separates a knowledge-level Claim Proposal from the resulting text diff, allowing reviewers to approve additions and adjudicate conflicts rather than silently overwriting facts. Evaluated on 599 English Wikipedia articles from MegaWika 2.0, spanning 49 source languages, KPR achieved MiRAGE information precision of 0.878, retained 0.950 of existing information, and added 0.891 of source information, outperforming raw-text ConText and unfiltered claim-based ConClaim. On multilingual question answering, KPR grounding averaged 69.2 accuracy versus 58.5 for ConClaim and 48.4 for ConText, while preserving English-content performance. On a difficult 100-question subset, GPT-5.6 answered 62 questions with a KPR-revised article, compared with 38 using web search; even smaller models such as Llama-3.1-8B benefited. KPR edits preserved 97.3 percent of the original Wikipedia text and required 11.2 contiguous approvals, compared with 25.3 for ConClaim. In RAGTIME report updating, the same claim-level process improved retention and addition across temporal, conflict, and balanced settings, while explicitly withholding contradictory claims. The pipeline used Qwen3.5-27B for decomposition, classification, routing, and rewriting, demonstrating that continual authoring can be both information-complete and auditable.