Evaluation and verification — limits of LLM judges and the rise of traceable evaluation

Verification-first evaluation is expanding from programmatic checks and research agents to realistic user-request and conflict-sensitive benchmarks. One-Eval, MM-CondChain, MiroThinker-1.7/H1, AgentProcessBench, PACE, and RealSWE emphasize traceability, hidden constraints, process quality, and realistic task distributions rather than relying solely on LLM judges or final-task success. Security concerns around agentic blabbering and phishing further raise the priority of auditable tool use.