ClawProBench Closed Dataset Release and Analysis
A bilingual release note for the 68-task closed dataset, explaining how closed enterprise workflows differ from the open leaderboard and why some model rankings move sharply.
Bench Blog
Updates, benchmark notes, result interpretations, and design changes for ClawProBench.
Safety Under Live Agent Work: What the ClawProBench Leaderboard Shows
A model-family analysis of safety scores, hard safety failures, and secret-refusal behavior across the current ClawProBench leaderboard.
My Feelings During the Development of ClawProBench
A bilingual note on why I built ClawProBench, how the harness was shaped through self-iteration, and what I learned from running different models and coding plans.
Open-sourcing ClawProBench: Bringing Agent Benchmarks Back to the Real Runtime
ClawProBench is designed to evaluate model intelligence under OpenClaw across planning, tool use, constraints, recovery, synthesis, and safety.