Evaluation Theater: How Structural Compliance Decouples from Cognitive Judgment in Deployed LLM Agents

Deployed LLM agent pipelines increasingly adopt structured-output safety gates—function-call validators, schema enforcers, evidence-completeness checks—as cheap, deterministic surrogates for harder-to-measure cognitive properties. We show this surrogate can fail completely. In a security-operations pipeline simulation across 8 frontier LLMs, an Evidence Addressal Completeness gate yields identical perfect scores (primary completeness = 1.0, gate pass rate = 100%) on both 144 legitimate verdict revisions and 49 social-engineering capitulations: a discriminative gap of 0.0 percentage points. Across the 193 changed-verdict trials the gate's blocking logic never once engaged. We name this failure mode evaluation theater, characterize it with a mirror-pair diagnostic, document a "recognize-but-capitulate" pattern in which models articulate suspicion in structured output before flipping verdicts anyway, and propose a Cognitive Compliance Index built on content signals the structural gate ignores (AUROC 0.934, 95% CI [0.899, 0.966]).