Legal AI, not a coding agent with scaffolding
4 hours ago
- #agent-workspace
- #evidence-grounding
- #legal-ai
- Legal AI systems should be built specifically for legal purposes, not repurposed coding agents.
- Key functions include finding supportable arguments, identifying evidence, and enabling verification through an agent workspace with audit trails and user control.
- Agent grounding must be granular at the claim/edit level, not just citing whole documents, with explicit source hierarchy and purpose (e.g., legal authority vs. client facts).
- AI suggestions should show reasoning (instructions, clauses, facts, legal sources, warnings) and present edits as tracked changes for user review, not automatic changes.
- Version control is essential to log and revoke changes, preventing reckless edits by probabilistic processes.
- Skills and memory should preserve exact evidence access via lookup artifacts, not replace it with summaries, to maintain factual integrity.
- AI must not subvert legal judgment; systems should highlight source limitations (e.g., outdated, jurisdiction-specific) to inform lawyer decisions.
- Deleting clauses requires checks for text changes and dependencies; systems like Codex and Lexifina use different methods (textual invariants vs. structured cross-reference reviews) to prevent errors.
- Compaction should allow recovery of exact tool outputs via mechanisms like tool_use_id lookups, avoiding reliance on paraphrases for evidence.
- Redlines should prove drafting decisions by tracing changes to specific versions, instructions, and edits, with systems like Codex providing net diffs and Lexifina offering provenance with confidence levels.