Benchmark Radar Day 102: A Verifier Without Evidence Is Just Another Opinion

Day 102 of Benchmark Radar. The scan fetched 1,638 items, published 785, and marked 354 as recommended.

Author: Koutian Wu; GitHub: ktwu01

Benchmark Radar is live. Open the dashboard, star the repository, or read the paper. The Hugging Face dataset provides the data.

FORMA-Bench tests whether a designated verifier in a multi-agent architecture can check declared constraints from different evidence arrangements. It supplies 2,910 observations across five operational domains, including data residency and delegated financial authority, with both compliant and violating scenarios. The question is whether the architecture gives the verifier enough evidence to decide, not merely whether the underlying model can state the policy correctly. Dataset for the paper “Architectural Verifiability in Agent-to-Agent Systems: Design and Evaluation of FORMAW.

SlackBench grounds collaborative tasks in conversations, direct messages, real code repositories, and potentially stale or conflicting evidence. ViBench 2.0 and ReScopeBench separately test successive feature requests and experience reuse as environments change. Single-task accuracy may miss failures in finding, reconciling, and retaining information across a project. SlackBench: Benchmarking Agents on Collaborative Projects Grounded in Real Code Repositories.

Why this matters

Policy gates only work when the architecture hands the verifier a decidable record. Workplace agents need the same treatment over time, tested on evolving workflows rather than isolated prompts.

Scan coverage

This is a keyword-filtered discovery feed, not a representative survey of AI research. The archived briefing used 181 of 785 captured evidence records. These findings summarize its cited releases; discovery dates do not necessarily equal original publication dates.

Data and briefing: 2026-11-05 archived snapshot.