ProofPathBench: A Benchmark for Verifiability-Aware Planning by Tool-Using Language Agents

ProofPathBench is an implemented research benchmark for whether tool-using language agents select matched plans that produce independent evidence of external-state outcomes. Version 0.0.1 contains 96 synthetic scenarios across eight domains, 384 task-treatment units, 768 deterministic no-fault plan executions, and nine seeded failure classes. The deposit includes a research preprint and an Apache-2.0 benchmark software artifact. It reports no provider-model behavioral results; fixtures and prospective power simulations are not empirical model results. A blinded human construct-validation protocol is included but has not been run. The work has not undergone peer review. Author: Vamshi Krishna Madhavan, Independent Researcher. Correspondence: [email protected]. Source repository: https://github.com/Vamshi0104/proofpathbench. Project website: https://vamshi0104.github.io/proofpathbench/.