Grab Bench: Evaluating AI on Grab-shaped production work

What this post added

Introduced Grab Bench, an evaluation harness for AI systems on production-like tasks. Designed to evaluate AI models on specific contracts like metric faithfulness, tool parameter discipline, and evidence grounding, rather than just general confidence. Implemented a system with task plugins, deterministic scorers/LLM judges, and a focus on making failure modes visible. Developed a split between teaching and certification artifacts for reproducibility and generalization testing. Established gates for evaluating the harness itself.