The report
MolLean applies Lean-checked state transitions to an LLM's molecular reasoning. The October 14 paper describes an executable representation for editing molecular structures, evaluating assertions and returning verification feedback through MolLean-MCP. A rejected operation leaves the earlier chemical state unchanged instead of quietly committing an invalid graph transition.
The study compares direct prompting, one-pass execution, whole-trace refinement and interactive verification on 8,800 matched examples covering twelve tasks. The researchers report stronger results for verified refinement than direct chain-of-thought across the tasks tested with Qwen3.6-27B. These are benchmark outcomes in molecular generation, editing and reactions, not measured laboratory discoveries.
The distinction that matters for broader agent design is between a valid step and a correct goal. A proof checker can confirm that a graph edit preserves a structural condition while the model is still optimizing the wrong property, selecting the wrong target, or relying on an unverified external measurement. The MolLean Case records 1,533 structurally sound yet incorrect outputs in a 3,300-example ChEBI20 analysis. It also distinguishes native proof obligations from externally supplied witnesses.
Mechanical verification therefore changes the evidence available about an agent's intermediate operations; it does not infer the user's intended objective by itself. A consequential workflow needs a separately specified goal, checked transitions and attributed external observations. Treating a chain of valid steps as a guarantee of the desired final result would claim more than this experiment establishes.