OpenProver ships Planner-Worker-Verifier architecture for Lean 4 ATP with automatic formal verification

The system maintains a bounded Whiteboard scratchpad and unbounded Repository of partial proofs. A Planner agent issues subgoals to stateless Worker instances that emit tactic blocks; the Verifier compiles each block inside Lean 4 and returns kernel-level success or error messages. Automatic verification replaces LLM-as-judge scoring, producing binary pass/fail labels on every candidate proof.

On ProofNet, the released baseline records 18.4 % pass rate under a fixed 32-step budget per theorem. Ablations show that adding the Repository lifts performance to 24.1 % while interactive steering by a human operator reaches 31.7 % within the same budget. These numbers are obtained by re-running the exact Lean 4 environment recorded in the repository commit hash 4f9e2c1.

Prior agentic provers such as Aletheia and ReProver report comparable headline percentages yet omit kernel-verified proof objects, making direct comparison impossible. OpenProver closes that gap by exposing the full Lean 4 environment and a reproducible evaluation harness. The design therefore supplies the first public substrate for measuring whether agent scaffolding improves verifiable mathematical reasoning rather than surface-level tactic prediction.

Next milestone is integration with Lean 4’s upcoming mathlib4 release scheduled for October 2026; success will be measured by zero-regression on the 1 234 theorems that currently compile under mathlib4 HEAD.