Releases

Four additions bring the catalog to 223 entries: a benchmark of research judgment built from human corrections to LLM research agents, a toolkit that checks agents' success claims against recorded evidence, a coding-agent safety benchmark, and a library that enforces action protocols at run time.

SciΠvot's first author submitted its entry, agent-claimcheck's author suggested it, and the list maintainer selected V³AST-Bench and Concordat. None of the four is affiliated with Audit Commons. The summaries follow each project's paper or repository. Catalog inclusion does not certify any method or result.

Three additions and one removal leave the catalog at 219 entries. The additions are a decision-record runtime, a finance-specific specification for tamper-evident trading records, and a benchmark for privacy leakage between agents; the Praesidia audit verifier is removed because its repository is no longer available.

The solvi author submitted its entry, and the people who proposed VCP and AgentLeak disclosed their involvement in those projects. None of the three is affiliated with Audit Commons. The summaries follow each project's repository or paper. Catalog inclusion does not certify any method or result.

One addition brings the catalog to 213 entries: an offline verifier for exported Praesidia decision-audit bundles, contributed with a disclosed vendor affiliation.

Audit Commons summarizes the linked documentation and its own pre-merge build and test review of the verifier. The contributor disclosed an affiliation with Praesidia. Catalog inclusion does not certify safety or assess Praesidia's hosted platform.

Four additions bring the catalog to 217 entries. All four papers are co-authored by Audit Commons maintainer Yue Zhao, and the CatchBench record now links its paper.

Audit Commons maintainer Yue Zhao co-authored all four added papers and CatchBench, and each carries the Maintainer project tag in the catalog. The summaries follow the papers' abstracts. Catalog inclusion does not certify any method or result.

Three additions bring the catalog to 212 entries, covering local execution records, reward-hacking benchmarks and data, and a macOS policy broker prototype.

Audit Commons summarizes the linked documentation and has not independently tested these tools or replicated the research results. Catalog inclusion does not certify safety.

Four additions and two corrections bring the catalog to 209 entries. The update covers behavior evaluation, execution records, human approval, and financial AI assurance.

The descriptions summarize linked documentation. Audit Commons has not independently tested these tools; catalog inclusion does not certify software or ratify a framework.