Legal Quants Weekly Digest — Week 17

Oscar LQ stole the show with live demos of multi-turn contract redlining (including redline-on-redline via OOXML manipulation), DPIA automation backed by ICO RAG, and company secretarial work — capabilities that well-funded legal tech incumbents still lack. An 8-year-old independently discovered Claude Code and built an educational app, sparking a rich thread on children, AI, and critical thinking. A member delivered legal advice as interactive HTML instead of Word, and multiple others revealed they'd done the same — signaling HTML as an emerging format for legal content delivery. On the technical side, Google's TurboQuant KV cache compression paper dominated AI Research (6x compression with claimed zero accuracy loss), agent "dreaming" architectures were compared across multiple builders, and a member shared a detailed 6-phase framework for refactoring AI-generated codebases. The Claude Code vs. Codex debate continued, with subagent model selection bugs frustrating users. Openclaw reliability struggles prompted one member to share a breathtakingly detailed personal AI architecture featuring overnight dream cycles, a persistent "soul" layer, recursive self-improvement loops, and a hybrid local/cloud inference stack.