World Model Benchmarks Need Receipts, Not Just Scores

HarnessEval-W is a useful reminder that agent benchmarks need traces, evidence trees, and failure families, not just leaderboard rows.