#18 · ML / AI methods

LLM agents in biology gain provenance, audit trails and verification gates

The dominant thread is LLM agents operated inside explicit verification and governance layers rather than trusted on model output alone. Plato-Bio adds provenance tracking, citation checks and publication gates to biology novelty screening. CogEEGAgent controls analysis release through deterministic gates on MNE-Python EEG workflows and uses held-out confirmation to curb false positives. MetaClaw maps requests to registered workflows, runs upstream FlowHub processing and isolated downstream analyses, and emphasizes auditability across metagenomic and multi-omics data. The FEV framework offers a complementary evaluation lens, auditing systems by function, evidence and validation, and maps 109 systems and 28 benchmarks across biology domains. SymPerturb, a virtual-perturbation framework checked only by internal simulations, sits adjacent to this thread rather than within it. What is new is that auditability and provenance are built around the agent itself. All items are preprints, so the claims remain framework-level and unreviewed.