The Performance Layer: Observation Context Produces Classifiable Output Differences in a Language Model

A controlled experiment demonstrating that a single sentence about observation context produces classifiable output differences in an instruction-tuned language model (gen-only AUROC 0.692, p=0.008) while surface vocabulary remains identical (BoW 0.999). A three-condition design disambiguates observation sensitivity from sycophancy. Exploratory persona analysis suggests entity-relevant features may reduce observation sensitivity, with a named confound and registered decomposition replication. Includes dual-use considerations and an Open Questions section as a first-class methodological contribution.