Evidence · DiggingBeagle record
MolLean: executable verification for molecular reasoning
Original MolLean paper describing Lean-checked molecular state transitions, external-witness separation and stateful MolLean-MCP feedback, compared on 8,800 examples across twelve tasks. Verified steps do not alone establish fidelity to the requested objective.
- Published
- Oct 14, 2026
- Source role
- primary disclosure
Evidence record
Original MolLean paper describing Lean-checked molecular state transitions, external-witness separation and stateful MolLean-MCP feedback, compared on 8,800 examples across twelve tasks. Verified steps do not alone establish fidelity to the requested objective.
Claim-level citations (3)
- supportsLean-checked molecular actions can remain wrong for the user's objective: MolLean commits accepted molecular state transitions and leaves state unchanged after rejected operations, using Lean-checked representations and assertions.
Sections 2.1 through 2.5, executable state and verification design
- supportsLean-checked molecular actions can remain wrong for the user's objective: The authors distinguish process soundness, task fidelity and answer correctness. In their 3,300-example ChEBI20 evaluation, 1,533 outputs were sound but incorrect.
Section 2.4, verification boundaries; Section 3.3, ChEBI20 analysis
- supportsLean-checked molecular actions can remain wrong for the user's objective: MolLean evaluates one-pass execution, whole-trace refinement and interactive MolLean-MCP reasoning on 8,800 matched examples spanning twelve tasks, reporting that whole-trace refinement improves on direct chain-of-thought across those tasks for Qwen3.6-27B.
Paper abstract, benchmark scale and comparison of one-pass, refinement and MolLean-MCP modes