LLMs hit a seven-constraint reliability cliff

A deterministic benchmark of 15 models, 36 constraint types, and 369,753 checks found that even the strongest model’s success at satisfying every instruction falls below 50% at seven simultaneous constraints, while a model that passes individual constraints about 41% of the time at eight constraints satisfies all eight only 5.7% of the time, according to the Constraint Saturation Evaluation paper. Structural constraints degraded twice as quickly as lexical ones, and planning or self-correction barely moved the threshold. Agent builders should treat each added schema, safety rule, and formatting demand as a compounding reliability tax, alongside the archive’s evidence that harness design can dominate agent outcomes.