--- license: cc-by-4.0 tags: [agents, tool-use, benchmark, reinforcement-learning, actionproof] --- # ActionProof-Bench Frozen task splits of ActionProof, a SQLite customer-support simulator in which tool acknowledgements can be false by design (phantom, partial, stale and timeout-after-commit writes). Each task holds the prompt, the initial database state, the fault schedule and the goal predicates the deterministic verifier checks. - `*.jsonl` at the top level: the registered v1 splits (train / val / test_iid / test_ood_fault / variants / banking). - `a4/` … `a8/`: the splits and pools of Amendments 4–8 (training pools A/B, held-out credit workflow, unrecoverable tasks, recoverable twins, depth-3, forced prefixes, the Amendment 8 task list). The generator, environment, verifier and the hashes that freeze every split (`results/*hashes*.json`) are in the [code repository](https://github.com/atharvgaur1845/actionproof-v2). Tasks are synthetic; no real customer data.