Benchmarks on Latent Variablehttps://latentvariable.ai/tags/benchmarks/Recent content in Benchmarks on Latent VariableHugoen-usSat, 14 Mar 2026 00:00:00 +0000Your Agent Passed the Test by Breaking Every Rulehttps://latentvariable.ai/posts/your-agent-passed-the-test-by-breaking-every-rule/Sat, 14 Mar 2026 00:00:00 +0000https://latentvariable.ai/posts/your-agent-passed-the-test-by-breaking-every-rule/A new paper introduces ‘Procedure-Aware Evaluation’ and reveals that 27–78% of benchmark successes conceal procedural violations. No model achieves more than 24% reliable compliance. The implications for anyone deploying agents are significant.